JSCC AI Notes banner

Author: JSCC

  • Cursor Pricing Breakdown 2026: Hidden API Costs for 4 Monthly Users

    Cursor Pricing Breakdown 2026: Hidden API Costs for 4 Monthly Users

    Developers evaluating Cursor after the October 12, 2026 shift to The X Subscription need the current rates, not the old Pro numbers. Plus costs $8 a month and runs only Cursor models (Grok 4.7, 4.6, 4.5, and Composer 2.5). Premium starts at $30 a month, opens third-party models such as Claude Opus 5.5 and Sonnet 5.5 at their API list prices paid from prepaid credit, and multiplies the weekly pool. For four users who mix Grok with Claude on agent work, the difference between staying on Plus and moving to Premium or Teams is the prepaid credit burn plus the Cursor Token Rate on Teams.

    This pricing breakdown walks the live plan table, the model rate cards inside Cursor, and an illustrative four-user monthly bill. Every dollar figure below comes from cursor.com/pricing or Models & Pricing, checked November 5, 2026. Where we multiply those rates into a monthly total, the result is labeled illustrative — not a measured usage sample.

    Infocard for Cursor Pricing Breakdown 2026: Plus $8 Grok-only, Premium $30 with third-party at API list, Cursor Token Rate $0.25 per million on Teams, illustrative costs for 4 monthly users
    Four numbers that frame Cursor costs in 2026: Plus $8, Premium $30, Token Rate $0.25/M on Teams, and a 4-user bill.

    Current Plans and What They Include

    The X Subscription is the only product new customers can buy. Existing legacy Pro ($20), Pro+, and Ultra ($200) plans continue unless the subscriber switches. Annual billing is available: Plus $80, Premium $300, Super $1,000, Ultra $2,000. All paid plans share a weekly usage pool for Cursor models across Grok and Cursor. Plus locks those models to fixed settings (Fast off, medium effort) and excludes third-party models, Auto, Cloud Agents, Automations, Bugbot, and Origin. Premium and higher add those features and open third-party models.

    PlanPriceIncluded usageCursor modelsThird-party models
    Free$0LimitedLimitedNo
    Plus$8/mo ($80/yr)Base weekly poolGrok 4.7/4.6/4.5 + Composer 2.5 onlyNo
    Premium$30/mo ($300/yr)4x PlusSame, higher volumeAPI list from prepaid credit
    Super$100/mo ($1,000/yr)15x PlusSameAPI list from prepaid credit
    Ultra$200/mo ($2,000/yr)50x PlusSameAPI list from prepaid credit
    Teams$30/$100/$200 per seatMatches seat tierSameAPI list + $0.25/M Token Rate on eligible
    Compiled from cursor.com/pricing and cursor.com/docs/models-and-pricing, checked November 5, 2026.

    Individual plans do not pay the Cursor Token Rate. Teams and Enterprise do on eligible third-party requests (input, output, and cached tokens), as documented on the Token Rate help page. Models whose cache-read price is strictly below $0.10 per million tokens across every tier are exempt, as are Cursor models and Cost mode.

    Bar chart of Cursor monthly plan prices in 2026: Free $0, Plus $8, Premium $30, Super $100, Ultra $200
    Seat prices alone: Plus $8, Premium $30, Super $100, Ultra $200 per month before any third-party prepaid credit.
    Bar chart of Cursor weekly usage pool multipliers: Plus 1x, Premium 4x, Super 15x, Ultra 50x
    The weekly Cursor-model pool scales 1x → 4x → 15x → 50x from Plus to Ultra; third-party tokens stay outside the pool.

    Methodology / What We Checked

    We read the live Cursor pricing page, the Models & Pricing documentation, the Cursor Token Rate help page, and the official rate cards for the models that appear in Cursor. Claude Opus 5.5 lists at $4 input / $20 output / $0.20 cache-read per million tokens on Cursor’s third-party table (matching Anthropic’s published Opus pricing). Claude Sonnet 5.5 is $2 / $10 / $0.10. Grok 4.7 is $2 input / $6 output / $0.50 cache-read per million tokens on Cursor’s model table (standard context; long-context and Fast modes bill higher multipliers on the same page). Composer 2.5 standard is $0.50 / $2.50 / $0.20 (Composer 2.5 docs).

    No hands-on token measurements were run; every rate is taken from the linked official page. Illustrative monthly totals below use a fixed roughly 2 million tokens per user per month (about 500,000 tokens a week) and are labeled as such. A broken xAI developer pricing URL that returned HTTP 404 was dropped; Grok rates in this post come from Cursor’s own model table, with docs.x.ai/docs/models kept only as an alternate model hub link.

    Model Rates That Drive the Hidden Bill

    The seat price is only the floor. Once a team leaves Plus, every Claude, GPT or Google ($GOOGL) Gemini token is prepaid credit at list price. Teams also add $0.25 per million tokens on eligible third-party requests. The rates that matter for the scenarios below:

    Model / feeInput / 1MOutput / 1MCache read / 1MNotes
    Composer 2.5 (Cursor)$0.50$2.50$0.20Draws from weekly pool first; top-ups at these rates
    Grok 4.7 (Cursor)$2.00$6.00$0.50Standard context; Fast / long-context multiply
    Claude Opus 5.5$4.00$20.00$0.20Third-party; prepaid credit from first request
    Claude Sonnet 5.5$2.00$10.00$0.10Third-party; at the $0.10 cache-read floor
    Cursor Token Rate (Teams)$0.25 on eligible tokensInput + output + cached; Cursor models / Cost mode exempt
    Per-million-token rates from cursor.com/docs/models-and-pricing and the Token Rate help page, checked November 5, 2026.

    Sonnet 5.5’s $0.10 cache-read sits exactly at the exemption threshold language (“strictly below $0.10”), so treat Token Rate eligibility for that model as something to confirm in your admin bill rather than assume away. Opus 5.5 at $0.20 cache-read clearly pays the rate on Teams.

    Detailed Cost Scenarios for 4 Users

    A team of four can stay on Plus if they accept only Cursor models. At $8 each that is $32 a month plus any top-ups once the base weekly pool is exhausted. Top-ups are prepaid credit valid 365 days and are charged at the Cursor model rates (for example Composer 2.5 $0.50 input / $2.50 output).

    If any of the four needs Claude Opus 5.5 for long-horizon agent work, the account must be Premium or higher. The weekly pool still covers Cursor models; every Claude token comes out of prepaid credit at list price. An illustrative 500,000-token week per user that is half Composer 2.5 and half Claude Opus 5.5 looks like this (numbers are arithmetic on official rates, not measured usage):

    • Composer portion (250k input + 50k output): roughly $0.25–$0.40 depending on cache hit rate.
    • Claude Opus 5.5 portion (250k input + 50k output, modest cache): roughly $1.50–$2.00.
    • Across four users that is an extra $6–$10 a week in prepaid credit on top of the $120 Premium seats ($30 × 4).
    • Super and Ultra simply enlarge the Cursor-model pool; they do not change the third-party rate.

    Teams seats add the Cursor Token Rate of $0.25 per million tokens on eligible third-party requests. On the same Claude volume that is another $0.075–$0.10 per user per week. Cost mode and Cursor models remain exempt.

    # Illustrative monthly estimate for one user (not a measured result)
    # Assumes 2M tokens/month, 60% Composer / 40% Claude Opus 5.5, 30% cache hits
    composer_in = 1_200_000 * 0.50 / 1_000_000  # $0.60
    composer_out = 200_000 * 2.50 / 1_000_000   # $0.50
    claude_in = 800_000 * (0.7 * 4.00 + 0.3 * 0.20) / 1_000_000  # ~$2.30
    claude_out = 200_000 * 20.00 / 1_000_000    # $4.00
    print(f"Prepaid credit burn: ${composer_in + composer_out + claude_in + claude_out:.2f}")
    # Output: ~$7.40 before any Token Rate

    Scale that illustrative $7.40 by four users and you get about $30 of prepaid credit on top of $120 in Premium seats — roughly $150 a month before tax. The same seats on Teams add a small Token Rate line on the Claude share only. Plus stays near $32 if nobody leaves the Grok/Composer lane.

    Stacked bar chart of illustrative monthly cost for 4 Cursor users: Plus seats $32, Premium seats $120 plus about $30 Claude prepaid credit, Teams Premium $120 plus credit plus Token Rate
    Illustrative 4-user month: Plus ~$32 seats; Premium ~$150 with Claude prepaid credit; Teams adds a small $0.25/M Token Rate line.

    Who Should Pick Which?

    • Plus: choose it if the team is content with Grok 4.7 and Composer 2.5 and wants the lowest fixed cost. The $8 price is real only while the work stays inside those models.
    • Premium: choose it when any user needs Claude, GPT, or Gemini inside Cursor. Budget the prepaid credit separately; the $30 seat does not include those tokens.
    • Super or Ultra: make sense only if the Cursor-model weekly pool is the bottleneck (15x or 50x Plus). They do not discount Claude or GPT.
    • Teams: choose it when you need shared administration, SSO, or pooled usage. Factor the $0.25/M Token Rate on eligible third-party traffic; it is small on light Claude use and material on heavy agent loops.
    • Legacy Pro holders: compare the exact remaining Other Models allowance before switching, because the switch to an X Subscription tier is permanent and may include a prorated refund (account and billing help).
    Decision tree diagram for choosing a Cursor plan for 4 users in 2026: need third-party models leads to Premium plus prepaid credit; Grok and Composer only leads to Plus; pool bottleneck leads to Super or Ultra; team admin adds Teams Token Rate check
    Start with third-party need, then pool size, then team admin — three questions route a 4-user team to Plus, Premium+, Super/Ultra or Teams.

    A practical test for four seats: run one real backlog task for a week on Plus with Composer only, then the same task on Premium with Opus 5.5, and compare the prepaid-credit line in the dashboard to the illustrative $7.40 math above.

    FAQ

    Does Plus include Claude or GPT?

    No. Plus is limited to Cursor models (Grok family and Composer 2.5) at fixed settings. Third-party models require Premium or higher and are billed at API list price from prepaid credit.

    What is the Cursor Token Rate?

    $0.25 per million tokens on eligible third-party model requests for Teams and Enterprise only. It applies to input, output, and cached tokens. Individual plans and Cursor models are exempt. Models with cache-read prices strictly below $0.10/M across every tier are also exempt.

    How do annual prices compare?

    Plus is $80/year, Premium $300, Super $1,000, Ultra $2,000. That is roughly 17% off the monthly rate. Prices exclude tax.

    Can I still buy the old $20 Pro plan?

    Not as a new customer. Legacy Pro, Pro+, and Ultra continue for existing subscribers. Switching to an X Subscription tier is permanent and may include a prorated refund.

    Official Sources

    Related Posts

    About the author: JSCC writes about AI tools, developer software and the costs behind them, with a focus on turning pricing pages and docs into numbers readers can act on.

    This article was drafted with AI assistance and fact-checked against official documentation by the author. Prices and features were checked on November 5, 2026 and may change.

    © 2026 JSCC AI Notes

  • 6 Best Free AI Coding Agents in 2026: Feature Limits & Local Model Support Compared

    6 Best Free AI Coding Agents in 2026: Feature Limits & Local Model Support Compared

    Developers who want AI coding help without a monthly subscription in 2026 can start with open-source agents that run free models or point at a local model on their own hardware. OpenCode and Goose give the most flexible zero-cost routes because the tools themselves cost nothing and both accept local models through Ollama. GitHub Copilot Free, from Microsoft ($MSFT)-owned GitHub, supplies 2,000 completions a month with limited chat and agent use under automatic model selection. Google ($GOOGL) Antigravity offers a weekly quota with unlimited Tab completions and Command requests. Claude Code has no free tier and starts on the Pro plan.

    These six agents, plus a fully local Ollama setup that works with four of them, cover the main free paths available as of November 4, 2026. All prices, limits and feature descriptions come from the vendors’ own documentation or pricing pages, linked at the end.

    Infocard for the 6 best free AI coding agents in 2026: 6 agents compared, $0 tool cost for OpenCode, Goose, Cline and Aider, 2,000 Copilot Free completions per month, unlimited local use with Ollama
    Four numbers that frame a zero-cost AI coding setup in 2026: six agents, $0 tools, 2,000 Copilot Free completions, unlimited local runs.

    Free AI Coding Agents Compared

    Here is how the six agents stack up on what each one gives you for free, whether it can run a local model, and what the tool itself costs.

    ToolFree limits (official)Local model supportLicense / cost of toolKey notes
    OpenCodeFree models included, no account needed; rotating limited-time free models (e.g. DeepSeek V4.1 Flash Free, GLM-5.3-Flash Free, Flint 40B Free); usage cap not publishedYes (auto-discovers Ollama, LM Studio and vLLM; any OpenAI-compatible server)MIT, $075+ providers; free-model data terms vary (some zero-retention, some used for improvement)
    GooseFully free; no tool-level quotaYes (Ollama and 15+ other providers)Apache-2.0, $0 (Agentic AI Foundation at the Linux Foundation)Desktop, CLI and API; 70+ documented MCP extensions
    ClineRotating free model promotions with a limited quota (Cline account required); tool itself freeYes (Ollama, LM Studio, Atomic Chat)Apache-2.0, $0Optional ClinePass $9.99/month for 2-5x rate limits on selected open models
    AiderTool free; no quota on the softwareYes (Ollama, LM Studio, OpenAI-compatible APIs)Apache-2.0, $0Strong git integration; commits each change automatically
    GitHub Copilot Free2,000 completions per month; limited AI Credits for chat and agent use (amount not published); auto model selection onlyNot documented for the Free planFree plan, no credit cardNo Copilot in the cloud (cloud agent) on Free; Copilot CLI included
    Google AntigravityWeekly quota (size not published); unlimited Tab completions and Command requestsNo (no bring-your-own-key or custom endpoint)Free individual planGemini 3.8 Flash on every plan; some third-party models on Free are removed Nov 18, 2026
    Compiled by JSCC AI Notes from each vendor’s docs and pricing pages, checked November 4, 2026.

    Read the table from left to right as a cost question. Four of the six (OpenCode, Goose, Cline, Aider) cost $0 as software, so your only bill is whatever model you connect, and that can be $0 with a local model. The two commercial free plans trade that flexibility for a hosted model with a cap: Copilot Free’s 2,000 completions are roughly 65 completions a day over a 31-day month, while Antigravity doesn’t publish its weekly number at all.

    Chart of published free-tier limits for AI coding agents in 2026: GitHub Copilot Free shows 2,000 completions per month, while Antigravity, OpenCode free models and Cline promotions are marked as not published and Ollama local use as unlimited
    Only GitHub publishes a hard number (2,000 completions a month); every other free cap is either unpublished or set by the model you connect.

    Methodology / What We Checked

    We reviewed the official websites, documentation and pricing pages for each tool on November 4, 2026. Primary sources include opencode.ai, goose-docs.ai, docs.cline.bot, aider.chat, github.com/features/copilot and docs.github.com, antigravity.google documentation, claude.com and ollama.com. We recorded only limits and features that appear on those pages. Free model lists rotate, so the specific free models available inside OpenCode or Cline can change without notice.

    We did not run hands-on coding sessions or independent benchmarks for this guide; any speed or quality claims in the wider market are vendor-reported and are not repeated here as measured results. Hardware guidance for local models is taken from Cline’s local models documentation, not from our own tests. Where a vendor does not publish a number, we say so instead of estimating one.

    Detailed Setup Steps for Zero-Cost Paths

    OpenCode with Free Models or Local Ollama

    OpenCode states it is 100% free to use and includes a set of free models that work without creating an account. Install with:

    curl -fsSL https://opencode.ai/v2/install | bash

    After installation, run opencode and type /models to see available models, then pick one labeled free. Free models are described as available “for a limited time.” Some free models (for example DeepSeek V4.1 Flash Free and GLM-5.3-Flash Free) follow a zero-retention policy according to OpenCode’s console models documentation; others, such as Flint 40B Free, note that data may be used to improve the model and is kept for 30 days. OpenCode does not publish a usage cap for its free models, and you can hide them entirely with OPENCODE_DISABLE_FREE_MODELS=1.

    To switch to a local model, start Ollama and pull a coding model that fits your hardware. OpenCode’s model docs say it automatically discovers models from Ollama at http://127.0.0.1:11434 (and from LM Studio and vLLM at their defaults), so you select it with the ollama provider ID:

    {
      "$schema": "https://opencode.ai/config.json",
      "model": "ollama/gemma3:4b"
    }

    The tool remains free; you pay only for electricity and hardware. Swap gemma3:4b for whichever coding model you pulled.

    Goose with Local Models

    Goose is Apache-2.0 licensed and governed under the Agentic AI Foundation at the Linux Foundation. The desktop app (macOS, Linux and Windows) and CLI cost nothing. Install the CLI with the one-liner published on the goose site:

    curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash

    Point Goose at an Ollama model. Because the agent runs on your machine, prompts and code stay local when you use a local model. Goose supports more than 70 documented MCP extensions for additional tools, and it can also act as an ACP server for Zed, JetBrains or VS Code.

    Cline in an Editor

    Cline’s extension and CLI are free under Apache-2.0. After installing the extension in VS Code or another supported editor, sign in if you want access to Cline’s rotating free model promotions; Cline’s free models page says they are available to any user with a Cline account. Free models carry a limited usage quota, and free usage may be used to improve the models. Once the quota ends you can continue with your own API key, pay-as-you-go Cline credits, or the optional ClinePass subscription at $9.99 per month for 2-5x the rate limits on selected open models. For a pure local path, configure Cline to use Ollama, LM Studio or Atomic Chat.

    Aider for Git-Centric Workflows

    Install Aider with:

    python -m pip install aider-install
    aider-install

    Then run aider inside a git repository. Aider is free and model-agnostic. Supply an API key for a free-tier provider or point it at a local Ollama server, as described in Aider’s Ollama guide. Every change is committed to git, so you can review and revert with standard git tools.

    Copilot Free and Antigravity

    GitHub Copilot Free is available without a credit card. Official documentation lists 2,000 inline completions per month plus a limited allowance of GitHub AI Credits for chat and agent use; the exact credit amount for Free is not published. Model selection is automatic only, and Copilot in the cloud requires a paid plan. Antigravity’s individual free plan provides a weekly quota (number not published) with unlimited Tab completions and Command requests. Check the current model list because Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120b are scheduled for removal on November 18, 2026.

    What It Costs If You Outgrow Free

    Free tiers run out. Here is the nearest paid step for each path, from the official pricing pages, so you know the ceiling before you commit to a workflow.

    PathNearest paid optionPriceWhat the upgrade adds
    GooseNone listed$0Nothing to buy; you only pay for any hosted model you connect
    ClineClinePass$9.99/month2-5x rate limits on curated open coding models
    OpenCodeOpenCode Go$10/monthLower-cost access to a set of hosted models, with 5-hour, weekly and monthly limits
    GitHub CopilotCopilot Pro$10/month1,500 monthly AI credits, model selection, unlimited completions
    Claude CodeClaude Pro$20/month ($17/month billed yearly)Claude Code access; Claude’s Free plan does not include it
    Prices from claude.com/pricing, github.com/features/copilot/plans, opencode.ai and docs.cline.bot, checked November 4, 2026.

    Every paid option except Claude Pro lands at about $10 a month, so the cheapest upgrade is roughly $120 a year. Claude Pro at $20 is double that, and the yearly $17 rate still adds up to $204. If you already own hardware that runs a local model well, the $0 Goose or Ollama row is the only one that never grows.

    Bar chart of the nearest paid upgrade for each free AI coding path in 2026: Goose $0, ClinePass $9.99, OpenCode Go $10, GitHub Copilot Pro $10 and Claude Pro $20 per month
    Most upgrades sit at about $10 a month; Claude Pro, the entry point for Claude Code, costs $20.

    Going Fully Local: How Much RAM You Need

    A local model removes per-token billing and vendor quotas. Ollama’s pricing FAQ states that running models on your own hardware is always unlimited, and the local runtime itself is free. The catch is memory. Cline publishes a simple RAM guide that applies whichever agent you point at Ollama:

    System RAMTypical local setup (Cline docs)
    16-32 GBSmall or quantized models
    32-64 GBMid-size coding models
    64 GB+Larger models and bigger context windows
    Source: Cline, Local models documentation, checked November 4, 2026.

    In practice, a 16 GB laptop keeps you in Cline’s small or quantized tier. Moving to 32-64 GB opens up mid-size coding models, and 64 GB+ is where longer contexts become comfortable. Cline also recommends enabling its compact prompt and keeping tasks focused, since a smaller context means faster local responses.

    Range chart of RAM needed for local coding models according to Cline's docs: 16 to 32 GB for small or quantized models, 32 to 64 GB for mid-size coding models, and 64 GB or more for larger models with bigger context
    RAM decides how far a $0 local setup can go: 16-32 GB for small models, 32-64 GB for mid-size coding models, 64 GB+ for larger ones.

    Who Should Pick Which?

    • OpenCode: choose it if you want a free terminal, desktop or IDE agent with rotating free models and easy local model support. It is the lowest-friction start for many developers.
    • Goose: pick it when you prefer a vendor-neutral, foundation-governed agent with strong MCP extensibility and you plan to run local models for privacy.
    • Cline: use it inside your existing editor when you want approval steps before every file edit and terminal command, and you are comfortable with a free rotating quota or your own keys.
    • Aider: select it for git-native pair programming where every AI edit becomes a normal commit you can review and revert.
    • Copilot Free: stay on it if you already live in VS Code and mainly need inline completions; the 2,000 monthly limit is published and predictable for light use.
    • Antigravity: try it if you want an agent-first interface with a free weekly quota and unlimited Tab completions.
    • Fully local with Ollama: pair Ollama with OpenCode, Goose, Cline or Aider when privacy or unlimited offline use matters more than frontier model quality. Size the model to your RAM using the table above.
    Decision flow diagram for choosing a free AI coding agent in 2026: offline or private leads to Ollama with OpenCode, Goose, Cline or Aider; free hosted models lead to OpenCode or Cline; light VS Code use leads to GitHub Copilot Free; agent-first app leads to Google Antigravity; git-first terminal work leads to Aider
    Four yes-or-no questions route you to the right $0 path, from fully local Ollama to Copilot Free’s 2,000 completions.

    If you start on Copilot Free and later move to a paid plan, model choice starts to drive the bill, which is where routing tools come in. The quick test: run one real task from your backlog on two of these paths for a week each, and keep the one whose free limit you didn’t hit.

    FAQ

    Is Claude Code free?

    No. Anthropic’s pricing page lists Claude Code under the Pro plan, not the Free plan. Pro costs $20 per month, or $17 per month when billed yearly.

    Can I run a coding agent fully offline?

    Yes. Install Ollama, pull an open coding model that fits your hardware, and point OpenCode, Goose, Cline or Aider at the local endpoint. The software is free and there is no per-token charge.

    Do the free models inside OpenCode stay free forever?

    No guarantee. OpenCode documents that free models are available “for a limited time” and the list rotates. Some carry specific data-retention or training terms. Check /models inside the tool for the current set.

    What happens when I hit a free limit?

    Copilot Free’s 2,000 completions are a monthly allowance. OpenCode doesn’t publish a cap for its free models. When Cline’s free quota runs out, you can switch to ClinePass, Cline credits or your own API key. Local Ollama use has no vendor-imposed token limit.

    Official Sources

    Related Posts

    About the author: JSCC writes about AI tools, developer software and the costs behind them, with a focus on turning pricing pages and docs into numbers readers can act on.

    This article was drafted with AI assistance and fact-checked against official documentation by the author. Prices and features were checked on November 4, 2026 and may change.

    © 2026 JSCC AI Notes

  • GitHub Copilot Fusion vs Picking a Model Yourself (2026): What 1 Agent Task Costs on 11 Models

    GitHub Copilot Fusion vs Picking a Model Yourself (2026): What 1 Agent Task Costs on 11 Models

    If you pay for GitHub Copilot and keep hopping between Claude, GPT and Gemini models in the picker, this comparison is for you. GitHub, owned by Microsoft ($MSFT), made Copilot Fusion generally available on October 28, 2026, and it now chooses and combines models for you at three effort levels. Working only from GitHub’s published per-token rates, one mid-size agent task in our cost model ranges from 2.1 AI credits on the cheapest lightweight model to 212 on GPT-6 Astra, roughly a 100x spread. The short version: Fusion Standard is the sensible default for most people, while picking a model yourself still wins when you know exactly which one a job needs and want a predictable bill.

    Infocard comparing GitHub Copilot Fusion with manual model picking in 2026: 3 effort levels, 30% discount through November 30, about 100x cost spread per agent task, 1 AI credit equals $0.01
    Four numbers that frame the choice between Copilot Fusion and a fixed model pick.

    Copilot Fusion vs Manual Model Picking: Side-by-Side

    Here is how the two approaches differ on the points that affect your bill and workflow, based on GitHub’s documentation and changelog.

    What mattersCopilot FusionPicking a model yourself
    Who chooses the modelFusion routes each prompt and may chain several modelsYou do, and it stays fixed until you switch
    Effort settingsStandard, High, Max (Fast is internal to VS Code inline chat)Per-model reasoning level where the model supports it
    How it’s billedEach model it calls, at that model’s standard rate; no Fusion surchargeThe one model you picked, at its standard rate
    Current promo30% off model costs on paid plans through Nov 30, 2026None
    Cache behaviorTries to keep the main thread on one model so cached tokens keep workingCache resets whenever you switch models mid-session
    Local modelsStandard tier can route to MAI-Code-1.1-Flash or Qwen3.8 Flash Next (experimental)You can select a local model directly
    Cost predictabilityLower on High/Max, since extra review calls are added when Fusion thinks they helpHigh: one rate card, one model
    Where it worksVS Code, Copilot CLI, GitHub Copilot app, JetBrains IDEsEvery Copilot surface with a model picker
    Compiled by JSCC AI Notes from GitHub Docs and the October 28, 2026 changelog.

    Methodology: How We Compared Them

    We did not run hands-on benchmarks for this piece, so you won’t find any speed or accuracy scores we measured ourselves. This is a spec, pricing and documentation comparison. We read the About GitHub Copilot Fusion page, the Models and pricing for GitHub Copilot rate tables, the individual billing and organization billing pages, and the launch changelog, all checked on November 3, 2026.

    To turn per-million-token prices into something you can picture, we defined one illustrative agent task: a bug fix spread over about ten turns in a mid-size repository. We assumed 80K fresh input tokens, 320K cached input tokens (context the model reuses from earlier turns) and 20K output tokens. Each model’s cost is simple arithmetic on GitHub’s list rates, converted at 1 AI credit = $0.01. We left out cache-write fees, which some OpenAI and Anthropic models charge on top, so real numbers on those models will run slightly higher.

    One limit matters a lot: GitHub does not publish which models Fusion will pick for a given prompt. The Fusion figures further down are therefore hypothetical routes we built from the documented patterns, not readings from a real session. Your own usage dashboard is the only place to see true numbers.

    What One Agent Task Costs on Each Model

    Applying the same token profile to eleven models in the Copilot pool gives the spread below. Prices are GitHub’s default tier (shorter contexts); several models charge more past 200K or 272K input tokens per request.

    ModelInput / 1MCached input / 1MOutput / 1MCredits for our task
    GPT-6 Luna$0.10$0.01$0.502.1
    Claude Haiku 5.5$0.10$0.01$0.502.1
    MAI-Code-1.1-Flash$0.20$0.02$1.204.6
    Gemini 3.8 Flash$0.75$0.075$3.7515.9
    GPT-6.1 Terra$1.00$0.05$5.0019.6
    Claude Sonnet 5.5$2.00$0.10$10.0039.2
    GPT-6.1 Sol$2.00$0.10$10.0039.2
    Gemini 4 Argon$2.00$0.10$10.0039.2
    Claude Opus 5.5$4.00$0.20$20.0078.4
    Claude Fable 5.1$10.00$0.25$50.00188.0
    GPT-6 Astra$10.00$1.00$50.00212.0
    Rates from GitHub’s Models and pricing page (default tier). Task: 80K fresh input, 320K cached input, 20K output.
    Horizontal bar chart of estimated GitHub Copilot AI credits for one agent task across 11 models, from 2.1 credits on GPT-6 Luna and Claude Haiku 5.5 to 212 credits on GPT-6 Astra
    Same work, very different bills: our illustrative task costs about 2 credits on lightweight models and over 180 on the top frontier tier.

    The lightweight group (GPT-6 Luna, Claude Haiku 5.5, MAI-Code-1.1-Flash) stays under 5 credits, so even a heavy month barely dents an allowance. The versatile and powerful middle, where Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon from Google ($GOOGL) all share a $2 input and $10 output rate, lands at 39.2 credits. GPT-6.1 Terra sits at half that, 19.6. At the top, Claude Opus 5.5 doubles the middle tier, and Fable 5.1 or GPT-6 Astra run about five times the middle tier.

    Output tokens drive more of the bill than most people expect. On Sonnet 5.5, our 20K output tokens cost 20 credits, more than the 80K fresh input tokens (16 credits). Agents that write long explanations or rewrite whole files burn credits fast. Anthropic’s pricing choices here tie into the wider economics of frontier labs we looked at in our breakdown of the Anthropic S-1 filing.

    The hidden cost of switching models mid-session

    GitHub’s cost optimization guide warns that a different model can’t reuse another model’s cache, and it notes cached tokens are typically billed at about 10% of the normal input price. Say your context has grown to 40K tokens and you jump to Sonnet 5.5 halfway through. That next turn re-sends the full 40K as fresh input: 8 credits instead of the 0.4 credits it would cost as a cache hit. Repeat that often and switching can cost more than the model choice itself. Fusion is designed to route along cache boundaries, which is one of the quieter reasons it can come out cheaper.

    How Fusion’s Three Patterns Change the Bill

    Fusion doesn’t just pick one model. Per prompt, it uses one of three execution patterns, and only the first behaves like a normal single-model request.

    • Single: one model handles the prompt. Fusion Standard picks one model per task, so this is what you’ll see most at that level.
    • Cascade: an efficient model writes a draft, a judge reviews it, and the work is either accepted or escalated to a stronger model.
    • Critique: one model drafts while several others critique it in parallel; their notes are merged and the original model revises.
    Flow diagram of GitHub Copilot Fusion execution patterns: Single with one model, Cascade with a draft, judge and optional escalation, and Critique with three parallel critics and a revision step
    Every box in this diagram is a separate model call, and each one is billed at its own model’s rate.

    Because cascade and critique add calls, the docs say a task on Fusion High or Max may use more AI credits than it would on Standard. To show what that can look like, we priced four hypothetical routes using the same task profile, with the 30% launch discount applied. The judge reads 20K tokens and writes 1K on Sonnet 5.5; each critic reads 30K and writes 2K on GPT-6.1 Terra; the revision turn is small.

    Hypothetical routeCredits (after 30% off)vs manual Sonnet 5.5 (39.2)
    Single on GPT-6.1 Terra13.7-65%
    Cascade, draft accepted17.2-56%
    Critique with two critics36.8-6%
    Cascade, escalated to Sonnet 5.544.7+14%
    Hypothetical routes built by JSCC AI Notes from GitHub’s documented patterns and list rates. Not measured.
    Bar chart of hypothetical GitHub Copilot Fusion route costs with the 30 percent discount: 13.7 credits single, 17.2 cascade accepted, 36.8 critique, 44.7 cascade escalated, versus 39.2 for manual Claude Sonnet 5.5 and 78.4 for manual Claude Opus 5.5
    With the launch discount, most hypothetical Fusion routes undercut a fixed Sonnet 5.5 pick; an escalated cascade is the one case that costs more.

    When Fusion settles a prompt with a mid-priced model, it beats a fixed Sonnet 5.5 pick by a wide margin. When it escalates, you pay for the discarded draft plus the stronger model, which edges past a straight Sonnet run even with the discount. After November 30 the discount ends, and an escalated cascade in our model would cost 63.8 credits. GitHub’s own claim is more upbeat: it says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task (changelog). That’s GitHub’s figure, not ours, but it matches our math: the savings come from keeping routine work off frontier models.

    How to Turn On Copilot Fusion

    Update your client first, since older builds won’t list Fusion.

    1. VS Code or JetBrains: open the Copilot Chat model picker, choose Copilot Fusion, then pick Standard, High or Max.
    2. GitHub Copilot app: click the model name under the prompt box, switch the Copilot Fusion toggle on and choose an effort level (app docs).
    3. Copilot CLI: type /model in a session and select Copilot Fusion, or pass it as a flag for one-off prompts (CLI docs). Use /config model to make it your default.

    For scripted runs, the documented flags look like this (the prompt text is our own example):

    copilot --model fusion --fusion-tier standard -p "Add input validation to src/signup.ts and update its tests"

    To sanity-check a bill before committing to a model, this short snippet of ours does the math; plug in GitHub’s rates and token counts from your usage dashboard.

    def copilot_credits(fresh, cached, output, rate_in, rate_cached, rate_out, discount=0.0):
        """Estimate GitHub AI credits (1 credit = $0.01). Rates are USD per 1M tokens."""
        usd = (fresh * rate_in + cached * rate_cached + output * rate_out) / 1_000_000
        return round(usd * 100 * (1 - discount), 1)
    # Claude Sonnet 5.5 at list rates: 39.2 credits for our sample task
    print(copilot_credits(80_000, 320_000, 20_000, 2.00, 0.10, 10.00))

    Local routing is a separate, experimental option. According to the local models changelog, on supported Windows machines you can install MAI-Code-1.1-Flash, and Fusion Standard may then send suitable work to it without using AI credits.

    How Far Each Plan’s AI Credits Go

    Since June 1, 2026, Copilot bills by usage. Each plan includes a monthly AI credit allowance that resets at 00:00 UTC on the first of the month, and unused credits don’t roll over. Code completions and next edit suggestions are not billed in credits and stay unlimited on paid plans.

    PlanPriceMonthly AI creditsTasks on TerraTasks on Sonnet 5.5Tasks on Opus 5.5
    Copilot Pro$10/mo1,500763819
    Copilot Businessper user1,900964824
    Copilot Enterpriseper user3,9001989949
    Copilot Pro+$39/mo7,00035717889
    Copilot Max$100/mo20,0001020510255
    Allowances from GitHub Docs. Pro, Pro+ and Max totals include base credits plus a flex allotment. Business and Enterprise credits are per user and pooled across the organization.
    Grouped bar chart showing how many illustrative agent tasks each GitHub Copilot plan's monthly AI credits cover on GPT-6.1 Terra, Claude Sonnet 5.5 and Claude Opus 5.5, from 19 Opus tasks on Pro to 1,020 Terra tasks on Max
    A $10 Pro plan covers about 38 Sonnet-class agent tasks a month in our model; the same credits stretch twice as far on GPT-6.1 Terra.

    Pro users feel model choice the most. Running every agent task on Opus 5.5 empties a Pro allowance in about 19 tasks, while routine work on Terra stretches it to 76. Copilot Free and Student plans only reach models through automatic selection, so the manual-versus-Fusion question mostly applies to paid tiers. If you hit the ceiling often, GitHub suggests that a bigger plan can work out cheaper than buying extra usage, and a $0 budget with a hard stop prevents surprise charges. Budget-conscious users can also look at the best free AI coding agents in 2026 if you want a free option.

    Who Should Pick Which?

    • Copilot Pro on a tight budget: use Fusion Standard as your default. It keeps the main thread on one model and can route to cheaper ones without you thinking about it.
    • You already know the right model: pick it manually and stay on it for the whole session. A fixed Sonnet 5.5 or Terra run gives you a predictable, easy-to-audit bill.
    • Hard bugs or multi-file refactors: try Fusion High for those sessions only. The review calls cost extra, but they’re cheaper than defaulting to a frontier model all day.
    • Architecture or deep debugging: Fusion Max or a manual Opus 5.5 pick. Plan with the strong model, then start a fresh session on a cheaper one to implement.
    • Team admins: set Fusion as the managed default and limit its model pool through policy. Fusion only uses models your org has enabled.
    • Windows developers with plenty of RAM: experiment with local routing on Fusion Standard. Check the hardware requirements first.

    Spending on AI tools rides on a much bigger buildout of chips and data centers, which we’ve tracked in our pieces on AI infrastructure debt and AI debt concerns in semiconductors. Falling list prices, such as Opus 5.5 pricing below Opus 5, help make routers like Fusion worthwhile.

    FAQ

    Does Copilot Fusion cost extra?

    No. GitHub says there is no separate charge for Fusion. You pay for each model it calls at that model’s standard rate, and through November 30, 2026, paid plans get 30% off those model costs.

    Is Fusion always cheaper than picking a model?

    Not always. Fusion Standard usually is for routine work, but Fusion High and Max can add judge or critic calls, so an escalated task can cost more than one run on a single mid-tier model.

    Can I see which models Fusion used?

    Yes. After a response, hover over the effort level name under it to see the list of models used. You can also follow each step’s model while it works.

    Do code completions use AI credits?

    No. Code completions and next edit suggestions are not billed in AI credits and remain unlimited on paid Copilot plans.

    Which effort level should I start with?

    Start with Fusion Standard for everyday prompts. Move to High for well-scoped but difficult tasks, and save Max for the hardest reasoning work.

    Official Sources

    Related Posts

    About the author: JSCC writes about AI tools, developer software and the costs behind them, with a focus on turning pricing pages and docs into numbers readers can act on.

    This article was drafted with AI assistance and fact-checked against official documentation by the author. Prices and features were checked on November 3, 2026 and may change.

    © 2026 JSCC AI Notes

  • Tax-Loss Harvesting in 2026: The $3,000 Limit, the 30-Day Wash-Sale Rule and an $11,000 Worked Example

    Tax-Loss Harvesting in 2026: The $3,000 Limit, the 30-Day Wash-Sale Rule and an $11,000 Worked Example

    A red number in your brokerage account doesn’t feel like good news. But before December 31, a losing position can actually do some work for you. Selling it on purpose, so the loss offsets gains you already took this year, is what people call tax-loss harvesting.

    This guide walks through it the way you’d do it at the kitchen table: one step at a time, with a running number table. We’ll use a single example the whole way, an $11,000 paper loss next to a $6,000 gain, so you can swap in your own figures. The rules come from IRS Topic No. 409 and Publication 550, checked November 3, 2026. It’s general information, not tax advice.

    Tax-loss harvesting 2026: $3,000 annual deduction limit against ordinary income, $1,500 if married filing separately, 30-day wash-sale window
    The three numbers that shape every harvest: the $3,000 yearly limit, the 30-day wash-sale window, and the carryover.

    The rules in one short paragraph

    Capital losses first cancel out capital gains. If losses are bigger than gains, you can use the extra loss to lower your other income, but only up to $3,000 a year ($1,500 if you’re married filing separately). Whatever is left over carries forward to later years. And if you buy the same or a “substantially identical” investment within 30 days before or after the sale, the wash-sale rule pushes the loss into the future instead.

    Meet the example investor

    Dana is a single filer. Earlier this year she sold Fund A, which she’d held for several years, for a $6,000 long-term gain. She also owns Stock B, bought for $25,000 and now worth $14,000. That’s an $11,000 unrealized loss. She plans to stay invested in the market either way.

    To keep the tax math simple, we’ll assume her long-term gains are taxed at 15% and her ordinary income at a 24% marginal rate. Your rates may differ; our capital gains tax guide explains how the 0%, 15% and 20% long-term rates work.

    Step-by-step: running the numbers

    Step 1: List the gains you’ve already realized

    Pull your realized gains from your broker’s year-to-date tax page. Keep short-term (held one year or less) and long-term (held more than one year) apart, because short-term gains are taxed like ordinary income. Dana has $6,000, all long-term.

    Step 2: Find positions with a loss you’d be okay selling

    Look for holdings where today’s value sits below what you paid. Dana’s Stock B shows an $11,000 loss. If it’s held more than a year, the loss is long-term too.

    Step 3: Net the loss against the gain

    $6,000 gain minus $11,000 loss leaves a $5,000 net capital loss. The $6,000 gain is now fully cancelled, so there’s no capital gains tax on it this year.

    Step 4: Apply the $3,000 limit

    Of the $5,000 net loss, Dana can use $3,000 against her wages and other income. That’s the lesser of $3,000 or her total net loss, as Topic 409 puts it. The loss goes on Schedule D and then onto her Form 1040.

    Step 5: Carry the rest forward

    The remaining $2,000 rolls into 2027. It can offset gains next year, or another slice of ordinary income if she has no gains. The IRS Capital Loss Carryover Worksheet in Publication 550 tracks it.

    StepWhat happensRunning total
    1Realized long-term gain (Fund A)+$6,000
    2Sell Stock B at a loss−$11,000
    3Net capital result−$5,000
    4Deduct against ordinary income (cap)−$3,000 used
    5Carryover to 2027−$2,000 left
    Illustrative example by JSCC Money Notes, applying the IRS $3,000 limit from Topic 409.
    Worked example bar chart: $6,000 long-term gain, $11,000 harvested loss, $5,000 net loss, $3,000 deducted, $2,000 carried forward to 2027
    Dana’s harvest in five bars. The gain disappears, $3,000 trims ordinary income, and $2,000 waits for next year.

    What it’s worth in dollars

    Here’s the estimated tax effect under our assumed rates. It’s a rough guide, not a promise, since your brackets, state tax and other income all change the picture.

    PieceAmountAssumed rateEstimated tax saved
    Long-term gain cancelled$6,00015%$900
    Ordinary income offset$3,00024%$720
    This year’s totalabout $1,620
    Carryover for 2027$2,000depends on next yearlater
    Rates are assumptions for illustration. Very high earners may also save the 3.8% surtax on investment income.

    If your income is high enough for the 3.8% Net Investment Income Tax, cancelling gains can trim that too. Our NIIT worked examples show where the thresholds start.

    The wash-sale rule: where most harvests go wrong

    Dana wants to stay in the market, so the natural move is to buy something right back. Publication 550 says a wash sale happens when you sell at a loss and, within 30 days before or after, you buy substantially identical stock or securities, get an option to buy them, or buy them in your IRA or Roth IRA. A purchase by your spouse counts too.

    The loss isn’t gone forever in most cases. It’s added to the cost of the new shares, so you get it back when you eventually sell those. The IRS example: buy 100 shares for $1,000, sell for $750, and rebuy within 30 days for $800. The $250 loss is disallowed now, and the new basis becomes $1,050.

    The IRA case is different. Publication 550 excludes it from that basis adjustment, so a repurchase inside an IRA can wipe out the benefit. If you’re also funding retirement accounts this year, our 401(k) vs IRA guide covers the order to fill them; just keep an eye on what those accounts buy around your sale date. For the bigger picture on where to park that money, see our guide to 401(k) and IRA contribution limits for 2026-2027.

    Wash-sale rule timeline: 30 days before and 30 days after a loss sale, with the IRS example of a $250 disallowed loss and $1,050 new basis
    Any substantially identical purchase in this window, including in an IRA, triggers the wash-sale rule (IRS Publication 550).

    A quick checklist before you sell

    • Check that the position is in a regular taxable account. Losses inside an IRA or 401(k) don’t count.
    • Look back 30 days. Did you, your spouse or an automatic dividend reinvestment buy the same holding?
    • Decide what you’ll hold for the next 31 days. Something that isn’t substantially identical keeps you invested without tripping the rule.
    • Sell before year-end, and confirm the trade shows on your broker’s realized gain/loss report.
    • Save the confirmation. You’ll report the sale on Form 8949 and Schedule D.

    Frequently asked questions

    Does the $3,000 limit cap how much gain I can offset?

    No. Losses first cancel gains without a cap. The $3,000 limit ($1,500 if married filing separately) applies only to the net loss left over that you use against other income, per IRS Topic 409.

    Can I harvest losses in an IRA or 401(k)?

    No. Publication 550 says its rules don’t apply to investments in IRAs, 401(k)s and other qualified plans, so there’s no deductible loss to harvest there.

    How long can I carry a capital loss forward?

    Topic 409 says you can carry the excess loss forward to later years. Use the Capital Loss Carryover Worksheet in Publication 550 or the Schedule D instructions to track it.

    Does the wash-sale rule apply if I buy before I sell?

    Yes. The window is 30 days before or after the sale, so buying extra shares in the month before a loss sale can also trigger it.

    Sources: IRS Topic No. 409, Capital Gains and Losses; IRS Publication 550, Investment Income and Expenses. Example figures are illustrative. This article is for general education and is not tax advice.

  • Altman’s Samsung Visit: 900,000 DRAM Wafers a Month and What It Means for PC Memory

    Altman’s Samsung Visit: 900,000 DRAM Wafers a Month and What It Means for PC Memory

    Quick answer: On October 31, 2026, OpenAI CEO Sam Altman met about 50 Samsung executives in Yongin, South Korea. The meeting was officially about AI transformation. Its backdrop is a memory deal from October 2025, when Samsung and SK hynix joined OpenAI’s Stargate project with a target of up to 900,000 DRAM wafer starts per month. That much AI demand is one reason PC memory has felt tighter.

    900,000 DRAM wafers per month: OpenAI Stargate memory target with Samsung and SK hynix

    What happened in Yongin on October 31, 2026?

    According to the Samsung Newsroom and Yonhap, Samsung hosted Altman at its Human Resources Development Center in Yongin. About 50 presidents from Samsung’s electronics, finance, and bio affiliates attended. The theme was “The next 10 years AI will change: choices for companies.”

    Samsung described the session as part of its company-wide AI transformation. It did not announce any new chip volumes. Yonhap separately cited industry sources saying Altman also met SK Chairman Chey Tae-won that day. Commentators guessed they may have talked about HBM supply, but no official details were released, so it’s best to treat that as speculation for now.

    Where does the 900,000 wafers number come from?

    It dates back to October 1, 2025. In its official announcement, OpenAI said Samsung Electronics and SK hynix plan to scale up advanced memory production, “targeting 900,000 DRAM wafer starts per month.” The companies described it in slightly different ways:

    • Samsung said OpenAI’s memory demand is “projected to reach up to 900,000 DRAM wafers per month,” and Samsung will help meet it.
    • SK Group said the figure is “more than double the current HBM industry capacity,” and that specific volumes will be decided in later talks.
    • Korea’s Ministry of Trade, Industry and Resources released a briefing on the same day’s meeting with President Lee Jae-myung.

    One thing to keep in mind: these were letters of intent, not fixed purchase orders. The 900,000 figure is an upper target for demand, not a confirmed shipment number.

    Timeline: Stargate memory LOIs October 2025 and Altman Samsung Yongin visit October 2026

    What does it mean for PC memory?

    Every wafer a memory maker gives to HBM and server DRAM for AI data centers is a wafer it can’t use for regular PC and phone DRAM. When one buyer signals demand on this scale, suppliers naturally put more of their capacity toward AI. For everyday buyers, that has usually shown up as firmer RAM and SSD prices and fewer deep discounts.

    A gentle takeaway: if you need a RAM or SSD upgrade soon, it may not pay to wait for big markdowns. If you don’t need one, there’s no need to rush either. Supply plans change, and nothing announced on October 31 changed the official numbers. To see where the wafers are actually going, check the latest numbers on Samsung and SK Hynix HBM revenue Q3 2026.

    For more background, see DDR4 vs DDR5 prices in 2026, Samsung and SK hynix HBM revenue, and our recent look at Kioxia’s NAND margins.

    FAQ

    Did Samsung announce a new deal with OpenAI on October 31, 2026?

    No. Samsung described it as an AI leadership seminar. No new volumes or contracts were announced.

    Is 900,000 wafers per month guaranteed?

    No. It is a target or projected demand level from the October 2025 letters of intent. SK said final volumes will be decided later.

    Does this affect investors?

    This post isn’t investment advice. If you’re thinking about taxes on chip-stock gains, our guides to capital gains tax and the 3.8% NIIT may help.

    About the author: JSCC follows the memory and AI infrastructure supply chain from Korea and reads the original Korean and English filings and newsroom posts behind each number. Every figure here links to its official source. Last checked November 1, 2026.

  • Net Investment Income Tax (3.8% NIIT) Explained: Thresholds and 4 Worked Examples

    Net Investment Income Tax (3.8% NIIT) Explained: Thresholds and 4 Worked Examples

    The Net Investment Income Tax (NIIT) is a 3.8% federal tax that quietly shows up for people whose income crosses a fixed line: $200,000 if you’re single, $250,000 if you’re married filing jointly. The tricky part isn’t the rate. It’s figuring out which dollars it applies to.

    Below we’ll walk through the rule in plain English, then run four worked examples so you can see exactly where your own numbers land. Figures are from IRS.gov, checked November 1, 2026. This is general information, not tax advice.

    Quick Summary

    • NIIT = 3.8% × the smaller of (a) your net investment income or (b) your MAGI above the threshold.
    • Thresholds: $200,000 single/head of household, $250,000 married filing jointly or qualifying surviving spouse, $125,000 married filing separately.
    • The thresholds are not indexed for inflation.
    • Wages don’t get taxed by NIIT, but they count toward MAGI and can push your investment income over the line.
    NIIT 3.8 percent thresholds by filing status: $200,000 single, $250,000 married filing jointly, $125,000 married filing separately
    NIIT thresholds by filing status. They are set by statute and not indexed for inflation (IRS).

    What is the Net Investment Income Tax?

    Since January 1, 2013, individuals have owed a 3.8% tax on the lesser of their net investment income or the amount by which their modified adjusted gross income (MAGI) exceeds a threshold based on filing status. That “lesser of” piece is the whole game, and it’s what the examples below are built around.

    Filing statusMAGI threshold
    Single$200,000
    Head of household$200,000
    Married filing jointly$250,000
    Qualifying surviving spouse$250,000
    Married filing separately$125,000
    Source: IRS, Net Investment Income Tax. Not indexed for inflation.

    What counts as net investment income?

    In general, net investment income includes interest, dividends, capital gains, rental and royalty income and non-qualified annuities, minus deductions properly allocated to that income. Taxable gain on investment real estate, including a second home, counts too.

    What doesn’t count: wages, unemployment compensation, Social Security benefits, alimony, tax-exempt interest, most self-employment income and operating income from a nonpassive business. Excluded gain on the sale of your principal residence is also left out.

    If you want a refresher on how the gains themselves are taxed before this extra 3.8% layer, our capital gains tax guide covers the 0%, 15% and 20% rates.

    How to calculate NIIT in three steps

    • Step 1: Find your MAGI and subtract your threshold. If the result is zero or less, you owe no NIIT. Stop here.
    • Step 2: Add up your net investment income.
    • Step 3: Take the smaller of Step 1 and Step 2, and multiply by 3.8%.

    For most U.S. filers MAGI is simply adjusted gross income; it’s increased by certain excluded foreign earned income if you live abroad (see the Form 8960 instructions).

    Four worked examples

    Example 1: Single, just under the line

    This one comes straight from the IRS Q&A: a single filer with $180,000 of wages and $15,000 of dividends and capital gains. MAGI is $195,000, which is below $200,000, so the NIIT is $0 even though there’s $15,000 of investment income.

    Example 2: Single, MAGI over by $30,000

    Wages $170,000 plus $60,000 of investment income gives MAGI of $230,000. Excess over $200,000 = $30,000. Investment income = $60,000. The smaller is $30,000, so NIIT = $30,000 × 3.8% = $1,140.

    Example 3: Married filing jointly, a big capital gain year

    A couple with $200,000 of wages sells a fund for a $100,000 gain. MAGI = $300,000, so the excess over $250,000 is $50,000. Investment income is $100,000. The smaller is $50,000: NIIT = $1,900. Only half of the gain was reached by the 3.8%.

    Example 4: Married filing jointly, high wages, small investment income

    $400,000 of wages and $20,000 of dividends: MAGI $420,000, excess $170,000. But investment income is only $20,000, and that’s the smaller number, so NIIT = $20,000 × 3.8% = $760. High wages alone never create NIIT; they only open the door.

    ExampleFiling statusMAGIExcess over thresholdNet investment incomeNIIT (3.8% × smaller)
    1Single$195,000$0$15,000$0
    2Single$230,000$30,000$60,000$1,140
    3Joint$300,000$50,000$100,000$1,900
    4Joint$420,000$170,000$20,000$760
    Example 1 is from the IRS Q&A; Examples 2–4 are illustrative math by JSCC Money Notes.
    Bar chart comparing MAGI excess and net investment income for four NIIT examples with resulting tax of $0, $1,140, $1,900 and $760
    The tax always follows the smaller bar.

    Gentle ways people manage the NIIT

    None of these are loopholes; they’re ordinary planning ideas worth discussing with a tax professional:

    • Watch the timing of big gains. Spreading a large sale across two tax years can keep more of it under the threshold.
    • Use pre-tax retirement contributions. Traditional 401(k) deferrals lower wages in AGI and therefore MAGI. Our 401(k) vs IRA guide explains the order to fill them.
    • Remember distributions. Distributions from IRAs and qualified retirement plans aren’t net investment income, but taxable ones do raise MAGI. Our RMD guide shows how big those can get.
    • Plan withholding. The NIIT isn’t withheld automatically. You can ask your employer for extra income tax withholding or make estimated payments to avoid a penalty.
    Checklist for NIIT: check MAGI, list investment income, take the lesser, multiply by 3.8 percent, file Form 8960
    A five-step checklist before you file Form 8960.

    NIIT vs. the 0.9% Additional Medicare Tax

    They started the same day and have similar thresholds, but they never hit the same dollar. The 0.9% Additional Medicare Tax applies to wages, compensation and self-employment income; the 3.8% NIIT applies to investment income. You can owe both in the same year, on different income.

    Frequently asked questions

    Are the NIIT thresholds adjusted for inflation?

    No. The IRS states the $200,000, $250,000 and $125,000 thresholds are not indexed for inflation, so more households cross them over time.

    Does the NIIT apply to my wages?

    No. Wages, Social Security benefits, unemployment compensation, alimony, tax-exempt interest and most self-employment income are not net investment income. Wages do count toward MAGI, though, which is what pushes you over the threshold.

    Is the gain on selling my home subject to the 3.8% tax?

    Only the part that is taxable for regular income tax. Gain you can exclude on a principal residence is not subject to NIIT; any taxable gain above the exclusion, or gain on a second home, can be.

    How do I report the NIIT?

    On Form 8960, filed with your Form 1040. If too little is withheld or paid in estimates, an estimated tax penalty can apply.

    Key takeaways

    • NIIT is 3.8% of the smaller of your net investment income or your MAGI above $200,000 (single) / $250,000 (joint) / $125,000 (separate).
    • The thresholds are not indexed, so it’s worth a quick check every year.
    • It’s reported on Form 8960 with your 1040.

    What this means for you: if your MAGI is within reach of the line, run the three-step math in December, while there’s still time to adjust the timing of a sale or a contribution. If you also own chip stocks, it helps to understand the AI debt concerns impacting Nasdaq semiconductors.

    About the author

    JSCC Money Notes is written by JSCC, an independent personal-finance writer who has spent years turning IRS publications into plain-English, calculator-style guides for everyday savers and investors. Every number in our articles is checked against IRS.gov or another official source on the date shown, and every worked example is recalculated by hand. We don’t sell products or give individual tax advice; for your own situation, please talk with a CPA or enrolled agent.

    Sources: IRS, Net Investment Income Tax; IRS, Questions and Answers on the Net Investment Income Tax; IRS Form 8960 (PDF); IRS Instructions for Form 8960 (PDF). Examples 2–4 are illustrative calculations by JSCC Money Notes and are not tax or investment advice. Checked November 1, 2026.

    © 2026 JSCC Money Notes

  • Kioxia’s 80% Operating Margin: NAND Joins the AI Memory Boom

    Kioxia’s 80% Operating Margin: NAND Joins the AI Memory Boom

    Kioxia, the Japanese NAND flash maker, earned an 80.2% operating margin in the July–September 2026 quarter on a non-GAAP basis (80.4% under IFRS). Revenue reached ¥2.48 trillion, up 453% from a year earlier. Together with Sandisk’s results a day earlier, it shows NAND, the memory inside SSDs, now earning margins that look a lot like the AI memory boom.

    Below is what the two reports actually say, how the margins are calculated, and what’s worth watching next. Figures come from the companies’ own releases, checked October 31, 2026.

    Executive Summary

    • Kioxia’s Jul–Sep revenue was ¥2.48 trillion (+453% YoY), with non-GAAP operating income of ¥1.988 trillion.
    • Operating margin: ¥1,988.3B ÷ ¥2,480.0B = 80.2% non-GAAP, or 80.4% under IFRS.
    • Next-quarter guidance is ¥2.5 trillion of revenue, with operating income guided to about ¥2.0 trillion.
    • Sandisk ($SNDK) posted $11.03B in revenue (+378% YoY), an 85.0% gross margin and a 79.6% GAAP operating margin.
    Kioxia 80.2% non-GAAP operating margin in the July to September 2026 quarter, with revenue of 2.48 trillion yen and operating income of 1.988 trillion yen
    Kioxia’s Jul–Sep 2026 quarter: ¥2.48T revenue, ¥1.988T non-GAAP operating income (Kioxia release, Oct 30, 2026).

    What did Kioxia report?

    For the July–September 2026 quarter, Kioxia reported revenue of ¥2,480.0 billion. That’s up 453.2% from the same quarter a year earlier.

    Operating income was ¥1,988.3 billion on a non-GAAP basis and ¥1,994.7 billion under IFRS. The company also declared an interim dividend of ¥90 per share.

    Item (Jul–Sep 2026)AmountMargin (formula)
    Revenue¥2,480.0B—
    Operating income, non-GAAP¥1,988.3B1,988.3 ÷ 2,480.0 = 80.2%
    Operating income, IFRS¥1,994.7B1,994.7 ÷ 2,480.0 = 80.4%
    Revenue growth, year over year+453.2%—
    Source: Kioxia Holdings earnings release, October 30, 2026. Margins calculated by JSCC Money Notes.

    Is 80.2% non-GAAP or IFRS?

    The 80.2% headline is the non-GAAP figure. Kioxia’s IFRS operating income was slightly higher, so the IFRS margin is 80.4%.

    The gap is small here, only 0.2 percentage points. Still, it’s worth naming the basis whenever you quote the number, since the two are measured differently.

    What does the ¥2.5 trillion guidance mean?

    This is the number most often misread. ¥2.5 trillion is Kioxia’s revenue guidance for October–December, not its profit guidance.

    Operating income is guided to about ¥2.0 trillion on a non-GAAP basis (¥1.995 trillion under IFRS). That implies a margin of roughly ¥2.0T ÷ ¥2.5T = 80%, close to this quarter’s level. The guidance assumes an exchange rate of ¥157 per dollar.

    How does Sandisk compare?

    Sandisk ($SNDK) reported its fiscal first quarter of 2027 on October 29, 2026. Revenue was $11,031 million, up 378% from $2,308 million a year earlier and up 23% from the prior quarter.

    Its GAAP gross margin was 85.0%, compared with 29.8% a year ago. GAAP operating income was $8,780 million, so the operating margin is $8,780M ÷ $11,031M = 79.6%.

    MetricKioxia (Jul–Sep 2026)Sandisk (fiscal Q1 2027)
    Revenue¥2,480.0B$11,031M
    Year-over-year growth+453.2%+378%
    Operating margin80.2% non-GAAP / 80.4% IFRS79.6% GAAP / 80.4% non-GAAP
    Next-quarter revenue guidance¥2.5T$12.2B–$12.6B
    Sources: Kioxia release (Oct 30, 2026); Sandisk release (Oct 29, 2026). Operating margins calculated by JSCC Money Notes.
    Bar chart comparing operating margins: Kioxia 80.2% non-GAAP and 80.4% IFRS, Sandisk 79.6% GAAP and 80.4% non-GAAP
    Two NAND makers, two accounting views, all within a point of 80%.

    Why is NAND suddenly so profitable?

    NAND flash has usually been the lower-margin, more cyclical corner of memory. These results suggest that has changed, at least for now.

    Sandisk’s segment numbers point to where the demand is coming from. Datacenter revenue rose 61% in one quarter, from $2,977 million to $4,802 million. CEO David Goeckeler described “outpaced growth in our data center business.”

    Sandisk segmentPrior quarterFiscal Q1 2027Change (QoQ)
    Datacenter$2,977M$4,802M+61%
    Edge$5,432M$5,832M+7%
    Consumer$556M$397M−29%
    Source: Sandisk fiscal Q1 2027 results, October 29, 2026.
    Sandisk revenue by segment: datacenter 4,802 million dollars up 61 percent, edge 5,832 million up 7 percent, consumer 397 million down 29 percent
    Datacenter SSDs drove Sandisk’s quarter while consumer sales fell.

    In plain terms, the growth is coming from data centers buying storage for AI workloads, not from people buying more USB drives. That’s the same pull that made DRAM and high-bandwidth memory so profitable. Our DDR4 vs DDR5 price breakdown shows how tight the DRAM side has become.

    What could change from here?

    Memory margins have swung hard in past cycles, so an 80% margin is a snapshot, not a promise. The clearest test is simple: do next quarter’s results match the guidance?

    For Kioxia, that means revenue near ¥2.5 trillion with operating income near ¥2.0 trillion. For Sandisk, it means revenue in the $12.2–$12.6 billion range.

    It also helps to watch how AI spending is being funded. We covered that in The AI Buildout Hits Its Funding Wall and in our look at what Anthropic’s S-1 actually says. Peers like Western Digital ($WDC) and Micron ($MU) are useful context when their own reports come out.

    Frequently asked questions

    Is Kioxia’s 80.2% margin GAAP or non-GAAP?

    It’s non-GAAP: ¥1,988.3 billion of operating income on ¥2,480.0 billion of revenue. On an IFRS basis the operating income was ¥1,994.7 billion, which works out to 80.4%.

    Is the ¥2.5 trillion figure profit guidance?

    No. ¥2.5 trillion is Kioxia’s revenue guidance for the October–December quarter. Its operating income guidance is about ¥2.0 trillion (non-GAAP), or roughly 80% of that revenue.

    How does Sandisk compare?

    Sandisk reported $11.03 billion of revenue, up 378% from a year earlier, with an 85.0% GAAP gross margin. Its GAAP operating income of $8,780 million equals a 79.6% operating margin.

    Will margins stay this high?

    Nobody can say for sure. Memory has a long history of boom-and-bust cycles, so the useful thing to watch is whether next quarter’s results line up with the guidance both companies gave.

    Key takeaways

    • Kioxia’s 80.2% is a non-GAAP operating margin (80.4% IFRS) on ¥2.48 trillion of revenue.
    • The ¥2.5 trillion figure is next-quarter revenue guidance; operating income is guided to about ¥2.0 trillion.
    • Sandisk’s 79.6% GAAP operating margin and 61% datacenter jump show the same AI-driven storage demand.

    What this means for you: when you see “80% margin” or “¥2.5 trillion” in a headline, check which basis and which line item it is, then compare next quarter’s results with the guidance.

    Sources: Kioxia Holdings news and earnings releases (Jul–Sep 2026 results, October 30, 2026); Sandisk fiscal Q1 2027 results (October 29, 2026). Margins are calculated by JSCC Money Notes. This is general information, not investment advice. Checked October 31, 2026.

    © 2026 JSCC Money Notes

  • Required Minimum Distributions (RMDs) in 2026: Age 73, the Formula and a $500,000 Example

    Required Minimum Distributions (RMDs) in 2026: Age 73, the Formula and a $500,000 Example

    Required minimum distributions (RMDs) start in the year you turn 73 for traditional, SEP and SIMPLE IRAs and most workplace plans such as 401(k), 403(b) and 457(b) accounts. The amount is simple: your balance on the prior December 31 divided by an IRS life-expectancy factor. At 73 that factor is 26.5, so a $500,000 IRA means a first RMD of about $18,868.

    This guide walks through when RMDs start, how to calculate one, the deadlines that trip people up and what a missed RMD costs. Figures are as of October 31, 2026, from IRS pages and Publication 590-B. It’s general information, not tax advice.

    Executive Summary

    • RMDs begin the year you reach age 73 for traditional IRAs and most workplace retirement plans.
    • Formula: prior December 31 balance ÷ Uniform Lifetime factor (26.5 at age 73, 24.6 at 75, 20.2 at 80).
    • Roth IRAs and designated Roth 401(k)/403(b) accounts have no RMDs while the owner is alive.
    • A missed RMD can cost a 25% excise tax on the shortfall, cut to 10% if corrected in time.
    Bar chart of IRS Uniform Lifetime Table factors from 26.5 at age 73 down to 12.2 at age 90
    The IRS factor shrinks as you age, so the share you must withdraw grows (IRS Pub 590-B, Table III).

    When do RMDs start?

    RMDs begin for the year you reach age 73. That covers traditional IRAs, SEP IRAs, SIMPLE IRAs and workplace plans, including 401(k), 403(b), 457(b) and profit-sharing plans.

    Your very first RMD gets a little grace period: you can delay it until April 1 of the year after you turn 73. Every RMD after that is due by December 31.

    Here’s the catch. If you use that delay, you take two RMDs in one tax year. Someone born in 1953 turns 73 in 2026, could take the first RMD as late as April 1, 2027, and would still owe the second by December 31, 2027.

    Timeline showing December 31 2026, April 1 2027 and December 31 2027 RMD deadlines for someone turning 73 in 2026, plus the 25% and 10% penalty
    Delaying the first RMD to April 1 stacks two taxable withdrawals into 2027.

    Still working? If you’re in your current employer’s plan and don’t own 5% or more of the business, you can generally delay RMDs from that plan until the year you retire. This exception does not apply to IRAs.

    How do you calculate your RMD?

    Divide each account’s balance on the prior December 31 by the factor for your age in IRS Publication 590-B. Most owners use Table III, the Uniform Lifetime Table.

    Use Table II instead if your spouse is your sole beneficiary and more than 10 years younger. Table I is for beneficiaries of inherited IRAs.

    AgeUniform Lifetime factorShare of balance (≈ 1 ÷ factor)
    7227.43.65%
    7326.53.77%
    7425.53.92%
    7524.64.07%
    7623.74.22%
    7722.94.37%
    7822.04.55%
    7921.14.74%
    8020.24.95%
    8516.06.25%
    9012.28.20%
    Source: IRS Publication 590-B, Appendix B, Table III. Percentages calculated by JSCC Money Notes.

    A $500,000 IRA at three ages

    The IRS’s own example is a 73-year-old with $34,800: $34,800 ÷ 26.5 = $1,313. The same formula on a bigger balance looks like this, holding the balance at $500,000 to keep the comparison clean.

    AgeFormulaRMD
    73$500,000 ÷ 26.5$18,867.92
    75$500,000 ÷ 24.6$20,325.20
    80$500,000 ÷ 20.2$24,752.48
    Example only. Real balances change every year, so recalculate with each December 31 value.
    Bar chart of RMDs on a $500,000 IRA: $18,867.92 at 73, $20,325.20 at 75 and $24,752.48 at 80
    Same balance, rising RMD: from $18,867.92 at 73 to $24,752.48 at 80.

    You can always withdraw more than the minimum. Just know the extra doesn’t count toward next year’s RMD.

    RMDs are taxed as ordinary income, except any part that’s a return of after-tax basis or a qualified Roth distribution. An RMD also can’t be rolled over into another tax-deferred account. If you want to see how that income lands in your bracket, our 2026 tax brackets guide lays out the rates.

    Which accounts need RMDs, and can you combine them?

    Account typeRMD while owner is alive?How to withdraw
    Traditional, SEP and SIMPLE IRAsYesCalculate each IRA, then take the total from any one or more IRAs
    403(b) contractsYesCalculate each contract, then take the total from any one or more 403(b)s
    401(k) and 457(b) plansYesTake each plan’s RMD from that plan
    Roth IRANoBeneficiaries follow RMD rules after the owner’s death
    Designated Roth 401(k)/403(b)NoBeneficiaries follow RMD rules after the owner’s death
    Source: IRS RMD FAQs (Q2, Q5).

    If you’re weighing where to save before retirement, our 401(k) vs IRA guide covers which account to fill first, and Roth dollars are the ones that skip lifetime RMDs.

    Charitably minded? A qualified charitable distribution (QCD) sent directly from your IRA to a qualified charity counts toward your RMD once you’re 70½ or older. The current Publication 590-B lists a $108,000 maximum annual exclusion, and the limit is indexed, so check the year you give.

    What happens if you miss an RMD?

    The amount you didn’t withdraw can face a 25% excise tax. It drops to 10% if you correct the shortfall in time, within two years. You report it on Form 5329.

    Example: a $10,000 shortfall × 25% = $2,500. Corrected in time, $10,000 × 10% = $1,000.

    The IRS can waive the tax if the shortfall came from a reasonable error and you’re fixing it. You file Form 5329 and attach a letter of explanation.

    Do Roth accounts have RMDs?

    Not while the owner is alive. Roth IRAs and designated Roth accounts in a 401(k) or 403(b) are exempt during your lifetime, but people who inherit them are subject to RMD rules.

    For most beneficiaries of owners who died after 2019, the whole account must be paid out within 10 years. Surviving spouses, minor children, disabled or chronically ill people, and beneficiaries not more than 10 years younger than the owner are the exceptions.

    How do states treat RMDs?

    Federal rules are the same everywhere, but state tax isn’t. States with no personal income tax, such as Florida, Texas and Nevada, don’t tax your RMD at the state level.

    Other states may tax it fully or only partly. Check your state revenue department before you set your withholding. Selling investments inside or outside the IRA? Our capital gains tax guide explains how taxable-account sales are treated differently.

    Frequently asked questions

    When is my first RMD due if I turn 73 in 2026?

    You can wait until April 1, 2027 for the first one. The second RMD is still due by December 31, 2027, so both would land in the same tax year.

    Can I delay RMDs while I’m still working?

    From your current employer’s plan, generally yes, until the year you retire, unless you own 5% or more of the business. The still-working delay does not apply to IRAs.

    Does a qualified charitable distribution count toward my RMD?

    Yes. A QCD sent directly from your IRA to a qualified charity counts toward your RMD. You must be at least 70½, and the current IRS Publication 590-B lists a $108,000 maximum annual exclusion; the limit is indexed, so check the current year.

    If I withdraw extra this year, does it reduce next year’s RMD?

    No. You can always take more than the minimum, but the IRS does not let the excess count toward a future year’s RMD.

    Key takeaways

    • RMDs start the year you turn 73; Roth IRAs and designated Roth accounts have none while you’re alive.
    • RMD = prior December 31 balance ÷ Table III factor (26.5 at 73). A $500,000 IRA means about $18,868.
    • Delaying the first RMD to April 1 means two taxable RMDs in one year, and a missed RMD can cost 25% of the shortfall (10% if fixed in time).

    What this means for you: put a calendar note for each December, check your December 31 balance against the factor for next year’s age, and you’ll rarely be surprised.

    Sources: IRS RMD FAQs; IRS Retirement Topics: RMDs; IRS Publication 590-B (Appendix B, Table III). Worked examples are calculated by JSCC Money Notes, are illustrative, and are not tax or investment advice. Checked October 31, 2026.

    © 2026 JSCC Money Notes

  • Anthropic S-1: $16.3B Revenue and an $80.7B Loss — What the Filing Actually Says

    Anthropic S-1: $16.3B Revenue and an $80.7B Loss — What the Filing Actually Says

    Anthropic, the company behind Claude, has filed to go public, and its headline loss looks alarming. The Form S-1 shows a net loss of $80.7 billion for the first six months of 2026. Revenue for the same period was $16.3 billion.

    Here is the part that is easy to miss. About $72.4 billion of that loss is a non-cash accounting charge tied to convertible notes, while the operating loss was just $693 million. The business itself came close to breaking even.

    All figures below come from the Anthropic, PBC Form S-1 on SEC EDGAR, filed on October 26, 2026. Half-year numbers are for the six months ended June 30 and are unaudited.

    Executive Summary

    • Revenue reached $16.3 billion in the first half of 2026, up from $1.2 billion.
    • The net loss was $80.7 billion, but $72.4 billion of it was a non-cash fair value charge.
    • The operating loss narrowed to $693 million, from $2.1 billion a year earlier.
    • Operating cash flow turned positive at $3.9 billion, and cash stood at $112.4 billion.

    What does the Anthropic S-1 say about revenue?

    Revenue grew very fast. The filing reports $16.286 billion for the first half of 2026, compared with $1.163 billion a year earlier.

    That is an increase of $15.1 billion, or 1,300%, in the company’s own words. Consumption-based revenue, where customers pay for what they use, added $12.1 billion of that gain.

    Subscription revenue added another $3.0 billion. For a sense of scale, Anthropic’s revenue for all of 2025 was $4.585 billion.

    Bar chart comparing Anthropic revenue of $1.2 billion in the first half of 2025 with $16.3 billion in the first half of 2026, and operating cash flow of negative $1.5 billion versus positive $3.9 billion
    Anthropic’s revenue grew 1,300% year over year in the first half of 2026, and operating cash flow turned positive. Source: Anthropic, PBC Form S-1, SEC EDGAR.

    Why is the Anthropic S-1 net loss $80.7 billion?

    Most of the loss comes from how the convertible notes are valued. These notes, many of them held by big cloud partners, can turn into Anthropic shares later.

    When Anthropic’s estimated value goes up, the notes become worth more on paper. Accounting rules record that increase as an expense, even though no cash leaves the company.

    The S-1 puts this change in fair value at $72.4 billion for the half year. It says the change was driven mainly by a rise in its enterprise valuation.

    • Operating loss: -$0.693 billion
    • Interest income: +$1.111 billion
    • Change in fair value of convertible notes, derivatives and SAFEs: -$72.388 billion
    • Other income (expense), net: -$8.757 billion
    • Loss before income taxes: -$80.727 billion
    • Provision for income taxes: $21 million
    • Net loss: -$80.748 billion

    You can check the math yourself. Start with the operating loss, add interest income, then subtract the two fair value lines and the small tax bill.

    Horizontal bar chart breaking down Anthropic's $80.7 billion net loss for the first half of 2026: a $0.693 billion operating loss, $1.111 billion of interest income, a $72.388 billion non-cash fair value charge on convertible notes, derivatives and SAFEs, and $8.757 billion of other expense
    Most of Anthropic’s first-half 2026 net loss came from a non-cash fair value charge on convertible notes, not from running the business. Source: Anthropic, PBC Form S-1, SEC EDGAR.

    How big was the Anthropic operating loss in 2026?

    The operating loss is the number that tracks the actual business. It was $693 million in the first half of 2026.

    A year earlier it was $2.135 billion, so the gap shrank even as spending grew. Total operating expenses were $16.979 billion, just above revenue.

    The largest cost was compute and infrastructure, at $8.945 billion. Personnel cost $3.152 billion, and sales, marketing and partnerships cost $3.168 billion.

    Anthropic also reports a non-GAAP figure called Adjusted Operating Income, which excludes stock-based pay and a few other items. It shows $1.618 billion for the half year, but it is the company’s own measure, not a GAAP result.

    How do the first halves of 2025 and 2026 compare?

    Metric (six months ended June 30)H1 2025H1 2026
    Revenue$1.163 billion$16.286 billion
    Total operating expenses$3.298 billion$16.979 billion
    Operating loss$2.135 billion$0.693 billion
    Fair value change on convertible notes, derivatives and SAFEs-$10.942 billion-$72.388 billion
    Net loss$12.969 billion$80.748 billion
    Net cash from operating activities-$1.513 billion+$3.929 billion
    Stock-based compensation$286 million$2.185 billion
    Source: Anthropic, PBC Form S-1 (SEC EDGAR), filed October 26, 2026. Half-year figures are unaudited.

    The table shows two stories at once. The operating side improved, and cash from operations swung from an outflow to an inflow.

    The accounting side moved the other way. The fair value charge grew as the company’s estimated value rose, which pushed the net loss much higher.

    How much cash and debt does Anthropic have?

    As of June 30, 2026, Anthropic held $112.4 billion in cash and cash equivalents. It also had $19.3 billion in short-term investments.

    The filing says the company has raised $141.1 billion in total capital so far. Its total indebtedness was $6.2 billion, which is the principal on convertible notes issued to Amazon ($AMZN) through Amazon Web Services.

    Alphabet ($GOOGL), through Google LLC, and Broadcom ($AVGO) have also signed note purchase agreements with Anthropic. The S-1 names Google as a direct competitor too, alongside OpenAI.

    Big compute deals are part of the picture. The S-1 says expected fixed or minimum guaranteed commitments under its infrastructure agreements were about $265.6 billion, with terms of one to 20 years. We looked at how this kind of spending is funded in our post on the AI buildout funding wall. For a smaller-scale comparison of AI compute spend, see what GitHub Copilot Fusion costs per agent task for a sense of scale.

    What should readers watch as the Anthropic IPO moves ahead?

    The S-1 is a first draft, and the offering price and share count are still blank. Numbers can change in later amendments, so check the newest filing on EDGAR.

    When you read headlines, look for whether they quote the net loss or the operating loss. They tell very different stories about the same six months.

    AI companies with large funding needs have moved markets before. Our coverage of CoreWeave’s bond yield and the Nasdaq’s financing-risk week gives useful background.

    Frequently asked questions

    How much revenue did Anthropic report in its S-1?

    Anthropic reported $16.3 billion of revenue for the six months ended June 30, 2026. That compares with $1.2 billion in the same period of 2025, according to the Form S-1.

    Why did Anthropic lose $80.7 billion?

    Most of the loss, about $72.4 billion, is a non-cash charge from the rising fair value of convertible notes, derivatives and SAFEs. The operating loss was only $693 million.

    Is Anthropic profitable?

    Not on a GAAP basis. It still had a small operating loss, but its operating cash flow was positive at $3.9 billion in the first half of 2026.

    How much cash does Anthropic have?

    The S-1 shows $112.4 billion in cash and cash equivalents as of June 30, 2026. It also held $19.3 billion in short-term investments.

    Which public companies are tied to Anthropic?

    The S-1 describes note agreements with Amazon, Google and Broadcom, and says Claude is offered on AWS, Google Cloud and Microsoft ($MSFT) Azure.

    Key Takeaways

    • Anthropic’s revenue jumped to $16.3 billion in the first half of 2026, up 1,300% from a year earlier.
    • About $72.4 billion of the $80.7 billion net loss is non-cash accounting on convertible notes.
    • The operating loss was $693 million, and operating cash flow turned positive at $3.9 billion.
  • How Capital Gains Tax Works in 2026: Short-Term vs Long-Term Rates

    How Capital Gains Tax Works in 2026: Short-Term vs Long-Term Rates

    Capital gains tax in the US depends first on how long you held the asset. Sell after one year or less and the gain is short-term, taxed as ordinary income at 10% to 37%. Hold it more than one year and the gain is long-term, taxed at 0%, 15% or 20% depending on your taxable income.

    For 2026, a single filer pays 0% on long-term gains while taxable income stays at or below $49,450, and 15% up to $545,500. This guide covers the 2026 thresholds, how gains stack on top of other income, two worked examples, the 3.8% surtax and how losses offset gains. Figures are as of October 30, 2026, from IRS publications. This is general information, not tax advice.

    Key takeaways

    • Short-term gains (held one year or less) are taxed as ordinary income at 10% to 37%.
    • Long-term gains (held more than one year) are taxed at 0%, 15% or 20% based on total taxable income.
    • Single filer with $60,000 of other taxable income: a $10,000 gain costs $2,200 short-term but $1,500 long-term, a $700 difference.
    • Net capital losses can offset up to $3,000 of other income per year ($1,500 married filing separately); the rest carries forward.

    How does the holding period decide short-term vs long-term?

    Per IRS Tax Topic 409, an asset held more than one year produces a long-term gain or loss. An asset held one year or less produces a short-term one.

    Count from the day after you acquired the asset, up to and including the day you sold it. Holding exactly one year still counts as short-term, so the sale has to land at least one day past the one-year mark.

    What are the 2026 long-term capital gains rates?

    Long-term gains are taxed at 0%, 15% or 20% based on taxable income (income after deductions). The 2026 thresholds come from section 3.03 of IRS Revenue Procedure 2025-32.

    Filing status0% rate up to15% rate up to20% rate above
    Single$49,450$545,500$545,500
    Married filing jointly / surviving spouse$98,900$613,700$613,700
    Head of household$66,200$579,600$579,600
    Married filing separately$49,450$306,850$306,850

    Because the thresholds use taxable income, the standard deduction comes off first. For 2026 it is $16,100 for single filers, $32,200 for married filing jointly and $24,150 for head of household (Rev. Proc. 2025-32, section 3.14).

    How are short-term gains taxed?

    Short-term net gains are added to your wages and other ordinary income and taxed at the regular graduated rates. For a single filer in 2026:

    RateTaxable income (single, 2026)
    10%Up to $12,400
    12%$12,401 to $50,400
    22%$50,401 to $105,700
    24%$105,701 to $201,775
    32%$201,776 to $256,225
    35%$256,226 to $640,600
    37%Over $640,600

    Want the full bracket tables and what is known about next year? See our 2027 tax brackets guide (official 2026 rates vs projected 2027).

    How much does holding past one year save? (Example A)

    Assume a single filer with $60,000 of ordinary taxable income (already after the standard deduction) who sells an asset for a $10,000 gain in 2026.

    • Long-term: other income already exceeds the $49,450 0% cap, so the whole gain is taxed at 15%: $10,000 × 0.15 = $1,500.
    • Short-term: taxable income goes from $60,000 to $70,000, all inside the 22% bracket ($50,400 to $105,700): $10,000 × 0.22 = $2,200.
    • Difference: $2,200 − $1,500 = $700 less federal tax by holding past one year.
    Bar chart comparing $2,200 short-term tax and $1,500 long-term tax on a $10,000 gain for a single filer in 2026
    Same $10,000 gain: $2,200 at the 22% ordinary rate versus $1,500 at the 15% long-term rate.

    How do gains stack on top of other income? (Example B)

    Long-term gains sit on top of your ordinary income. Whatever room is left under the 0% cap is taxed at 0%; the rest moves to 15%.

    Assume a single filer with $30,000 of ordinary taxable income and a $25,000 long-term gain, so total taxable income is $55,000.

    • Room left under the 0% cap: $49,450 − $30,000 = $19,450, taxed at 0%.
    • Remaining gain: $25,000 − $19,450 = $5,550, taxed at 15% = $832.50.
    • Effective rate on the whole gain: $832.50 ÷ $25,000 = 3.33%.
    Stacked bar showing $30,000 ordinary income, $19,450 of gain at 0% and $5,550 at 15% against the $49,450 cap
    Ordinary income fills the bottom first; only the part of the gain above $49,450 is taxed at 15%.

    Which gains are taxed above 20%?

    IRS Tax Topic 409 lists exceptions. Gains on collectibles (such as coins or art) and the taxable part of a section 1202 qualified small business stock gain are taxed at a maximum 28%. Unrecaptured section 1250 gain on real property is taxed at a maximum 25%.

    When does the 3.8% Net Investment Income Tax apply?

    On top of the regular rates, the Net Investment Income Tax adds 3.8% on the lesser of your net investment income or the amount your modified AGI exceeds these thresholds (IRS Tax Topic 559). Capital gains count as investment income.

    Filing statusNIIT threshold (MAGI)
    Single or head of household$200,000
    Married filing jointly / surviving spouse$250,000
    Married filing separately$125,000

    Unlike the capital gains brackets, these thresholds are not indexed for inflation.

    How do capital losses reduce the tax?

    Losses first offset gains. If losses exceed gains, you can deduct up to $3,000 of the net loss against other income each year ($1,500 if married filing separately), and carry the rest forward to future years. Losses on personal-use property, such as your home or car, are not deductible. For more, see tax-loss harvesting and wash sale rules before year-end.

    Sales are reported on Form 8949 and summarized on Schedule D (Form 1040).

    Do states and other countries tax gains the same way?

    No. Everything above is US federal tax. Many US states tax capital gains as ordinary income, and some states have no income tax at all. Other countries run their own systems, such as capital gains tax in the UK, so check your local rules.

    Tax-advantaged accounts change the picture too, since gains inside them are not taxed each year. See 401(k) vs IRA in 2026: which to fill first and how compound interest works for why untaxed growth adds up over time.

    FAQ

    How is the holding period counted for capital gains?

    Count from the day after you acquired the asset up to and including the day you sold it. More than one year is long-term; one year or less is short-term.

    What are the 2026 long-term capital gains rates for a single filer?

    0% on taxable income up to $49,450, 15% up to $545,500, and 20% above $545,500 (IRS Rev. Proc. 2025-32).

    Are short-term capital gains taxed differently?

    Yes. Net short-term gains are taxed as ordinary income at the regular 10% to 37% rates.

    Is the 3.8% Net Investment Income Tax adjusted for inflation?

    No. The thresholds stay at $200,000 for single or head of household, $250,000 for married filing jointly and $125,000 for married filing separately.

    How much of a capital loss can I deduct?

    Up to $3,000 of net capital loss per year against other income ($1,500 if married filing separately). Any excess carries forward.

    Sources

    IRS Revenue Procedure 2025-32 (sections 3.01, 3.03, 3.14; tax year 2026); IRS Tax Topic 409; IRS Tax Topic 559. Worked examples are calculated by JSCC Money Notes from these figures, exclude state taxes, and are illustrative, not tax or investment advice.