LAST_VERIFIED 2026-07-31
Context windows expand every few months; treat the table below as the current snapshot, not a permanent fact. Verify current limits against the vendor's own docs before sizing a production system around them.
TL;DR
The context window printed on a model's spec sheet is the number of tokens it accepts, not the number it uses reliably. Claude, GPT, and Gemini all publish large advertised windows, and all three degrade in accuracy well before that ceiling, a separate finding covered in depth in context rot. This post compares the advertised sizes model by model, and then the more useful number: roughly where quality starts to slip, so you can plan around the window that actually holds up rather than the one on the box. It's the budget branch's reference table for the context engineering series.
Sizing a system around a model's context window and want the real number, not the marketing one? We design AI agents around measured, tested windows, not advertised ceilings.
Key Takeaways
- Advertised context windows keep growing, but the token count a model handles reliably is consistently smaller than the number on the spec sheet, sometimes by a wide margin.
- Claude, GPT, and Gemini each publish different maximum windows across their model tiers; check the specific model and plan, not the vendor's largest headline number.
- Context rot means every model tested degrades as input grows, well inside its advertised limit, so a bigger window is not automatically a safer one.
- Claude Code and coding-focused tools often report a different effective window than the base model's API limit, because tool definitions and file context consume a meaningful share of it before a single line of the actual task arrives.
- This comparison is a living reference. Vendors ship new limits often enough that a static snapshot goes stale within months.
Context window sizes, compared
| Model family | Advertised context window | Notes |
|---|---|---|
| Claude (Sonnet, Opus, current generation) | Up to 1M tokens on supported tiers | Anthropic's docs list the exact limit per model and plan; 200K is the more common default tier |
| GPT-5 (API) | 400K tokens | Input and output share the same budget; check current OpenAI documentation for the exact split |
| Gemini (2.5 Pro and later) | Up to 1M-2M tokens on supported tiers | Google has offered the largest headline windows in the category; usable accuracy at the top end is the open question, not the ceiling |
| Claude Code / coding agents | Reported window minus tool and file overhead | The advertised model limit and the window actually available for your task differ once tool definitions and repository context are loaded |
Treat every row as the current best-known figure, not a permanent spec. Model tiers, plan-based limits, and beta windows change independent of a major version bump, which is exactly why this page is a living comparison and not a one-time snapshot.
The advertised number and the usable number are not the same
A model accepting a million tokens and a model answering reliably across a million tokens are different claims, and vendors advertise the first one. Chroma's research tested 18 models across major families and found every one degraded as input grew, well inside the advertised ceiling, on tasks that resemble real work rather than the easy version long-context benchmarks favor. A window that accepts a million tokens is not a claim that the millionth token gets read as carefully as the first.
The practical implication: size a system to the window where accuracy actually holds for your task, verified on your own data, not to the vendor's largest published number. Context rot covers how to test your own system for exactly this gap, with a focused-versus-full comparison that takes about an afternoon to run.
Claude context window, specifically
Claude's context window varies by model and plan; Anthropic's own documentation is the authoritative source for the exact current limit on the tier you're using, since it changes across releases and beta rollouts more often than a single number can capture. Claude Code and other coding-focused surfaces report a window that accounts for tool definitions and loaded files, which is smaller than the base model's raw API limit once that overhead is subtracted.
GPT-5 context window, specifically
GPT-5's API context window sits in the hundreds-of-thousands-of-tokens range, shared between input and output rather than allocated separately, which matters for tasks that need a long response as well as a long input. Check OpenAI's current documentation for the exact figure on the specific model variant in use, since the API and ChatGPT-facing limits are not always identical.
Gemini context window, specifically
Gemini has consistently published the largest headline windows in the category, into the millions of tokens on its top-tier models. The open question at that scale is less "does it accept the input" and more "does accuracy hold across it," the same context rot question that applies to every vendor in this comparison, not a Gemini-specific concern.
Frequently asked questions
What is Claude's context window?
Claude's context window depends on the specific model and plan; Anthropic's documentation lists the current figure per tier, with up to 1M tokens available on supported models and 200K as the more common default. Check the docs for the exact model in use, since limits change across releases.
What is GPT-5's context window?
GPT-5's API context window is in the hundreds of thousands of tokens, shared between input and output rather than counted separately. Confirm the exact current figure in OpenAI's documentation for the specific model variant, since API and consumer-facing limits can differ.
Which AI model has the largest context window?
Headline size leadership shifts as vendors ship updates, and Gemini has generally published the largest windows in the category, into the millions of tokens on top-tier models. A larger advertised window is not the same claim as reliable accuracy across that full window; verify both before choosing based on size alone.
Does a bigger context window mean better accuracy?
No. Research testing 18 models across vendors found every one degraded in accuracy as input grew, well inside its advertised limit. A larger window raises the ceiling on what a model accepts; it does not raise the reliability of what the model does with everything it is given.
Why is my Claude Code context window smaller than the advertised model limit?
Coding-focused tools load tool definitions, file contents, and repository context before your actual task begins, all of which consume tokens from the same budget as the base model's advertised window. The number available for your specific task is the advertised limit minus that overhead, not the full published figure.
