Context window is the spec everyone quotes. It is in every comparison table, every tool roundup, every Reddit thread about which assistant to pay for. It is also the spec almost nobody checks against the vendor, because it looks like the kind of number that stays still.
It does not stay still. On the morning of 10 November 2026 I corrected seven posts on this site that said Claude had a flat 200,000 token context window. By that afternoon I had corrected the same seven posts again, because I found a third Anthropic page that disagreed with the two I had used.
That second correction is what this article is about. If a vendor can give you three different answers in one day, the number in your comparison table is decoration.
So I read eight vendors' own pages and wrote down exactly what each one says.
What I did
For each vendor I read the pricing page first, the way a buyer would, then the model or API documentation. I recorded the exact wording, the model names, and the HTTP status of every page. Where a page refused an automated request I noted that and moved on rather than working around it.
Everything below was read on 10 November 2026. Nothing here comes from a third-party article, including the ones currently ranking for these queries.
Anthropic answers its own question three times
Three Anthropic pages, three levels of detail.
The pricing page has a Context window row. It reads "Up to 1M, varies by model". It reads exactly that for Free, Pro, Max, Ultra, Team and both Enterprise tiers. Every plan, same cell.
The usage and length limits article says: "The context window size depends on which model you're using. On paid plans, the newest models support up to a 1M token context window, while others support 500K or 200K tokens."
The per-model article gives the actual ladder. Chatting with Claude on a paid plan:
| Window | Models |
|---|---|
| 1M | Fable 5.2, Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5 |
| 500K | Fable 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6 |
| 200K | Everything outside that list |
None of these three pages is wrong. The pricing page gives a ceiling, the limits article gives a shape, and the per-model article gives the number. But only one of them answers the question a buyer is actually asking, and it is the one furthest from the checkout.
Then it changes again depending which product you open
This is the part I have not seen reported anywhere, and it is the most useful thing in this article.
The per-model article does not publish one table. It publishes three, one each for chatting in the Claude app, for Claude Code, and for Cowork. The same model gets different windows in each.
| Model | Claude app | Claude Code | Cowork |
|---|---|---|---|
| Opus 4.6 | 500K | 1M | 200K |
| Sonnet 4.6 | 500K | 1M | 200K |
| Haiku 5.5 | 1M | 1M | 500K |
| Opus 5.5 | 1M | 1M | 1M |
Read the Opus 4.6 row again. Same model. Same subscription. A five times difference between the best case and the worst, decided by which application you happened to open.
In Claude Code every model on the list is 1M, with a caveat worth knowing: Opus 4.6 and Sonnet 4.6 need you to select a specific variant with the /model command, and on Pro you have to enable usage credits to reach it at all.
If you are choosing a plan because of context, the surface you work in matters as much as the model you pick. No comparison table I have ever seen has a column for that.
The free plan appears in one answer and vanishes from the other
The pricing page shows "Up to 1M, varies by model" in the Free column, identical to every paid column.
The per-model article opens by scoping itself: it explains the window "on paid Claude plans (Pro, Max, Ultra, Team, Enterprise)". The free tier is not in the intro, not in any of the three tables, and not mentioned anywhere else on the page.
So Anthropic publishes a ceiling that includes free users and a detailed breakdown that excludes them. There is no published per-model figure specific to a free account. That is not a contradiction exactly, but if you are on the free plan and want to know what you actually get, the answer is that nobody has told you.
I got this wrong this morning. One of the seven corrections I made said the free tier gets the same figure as paid plans, because the pricing page shows that. It does show that. It is also the only page that does.
xAI ships newer models with smaller windows
xAI's model documentation has a column headed Context. The page never uses the phrase context window.
| Model | Context |
|---|---|
| grok-4.7 | 500k |
| grok-4.6 | 500k |
| grok-4.5 | 500k |
| grok-4.3 | 1M |
| grok-4.20 family | 1M |
| grok-build-0.1 | 256k |
The newest flagship has roughly half the context of grok-4.3. Whatever the engineering reason, a user upgrading from 4.3 to 4.7 on the assumption that newer means better would be downgrading on the one spec they probably checked.
I could not compare this against the consumer-facing page. x.ai/api returned 403 to an automated request, so the only context figures I can report for xAI are API ones.
Mistral publishes other companies' numbers but not its own
Mistral's models overview lists context windows for the third-party models it hosts:
- Z.ai GLM 5.3: 1M
- Z.ai GLM 5.3 Flash: 1M
- DeepSeek V4.1 Flash: 1M
- Mistral Moderation 2: 128k
Mistral Large 4, Mistral Small 4, Mistral Medium 3.5 and Ministral 3 carry no figure on that page. The individual model cards are separate pages and may well have one. But the overview page is where a person comparing models goes, and on that page Mistral tells you the context window of a DeepSeek model and not of its own flagship.
The pricing page does not state a context window anywhere. Its feature rows cover messages and web searches. The only context-adjacent text is a line about Project folders being "context-rich workspaces", which is marketing rather than a measurement.
What the rest publish
OpenAI. Both openai.com/chatgpt/pricing and the ChatGPT Plus help article returned 403. The developer documentation lists GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna each at "Context window 1.05M". Consumer plans unverified.
Google. The subscriptions page states a figure exactly once, inside the answer to "What is Google AI Pro?", which promises "a 1 million token context window". Free, AI Plus and AI Ultra have no figure in the portion of the page I could read. The developer docs give Gemini 4 Argon a 1M context. The subscriptions page never names which model its 1 million belongs to.
Microsoft. The plans and pricing URL returned 404. The Copilot hub on learn.microsoft.com loaded and states no context window at all, names no models, and gives no figures.
Perplexity. The FAQ hub returned 404. The API model docs state a window for exactly one model, Nemotron 3.5 Super at 1,048,576 tokens. Every other model on that page has none. The page also warns that a "context-length range" in its pricing table is a price threshold rather than a window, which is a distinction worth stealing.
DeepSeek. The only vendor in this set where one page answers the question. Pricing and documentation are the same page, and the model table has a row reading CONTEXT LENGTH with the value 1M for deepseek-flash and deepseek-v4-pro.
The scoreboard
| Vendor | Figure on the pricing page? | Figure in the docs? | Consumer plan answered? |
|---|---|---|---|
| Anthropic | Yes, a ceiling | Yes, full per-model ladder | Paid yes, free no |
| DeepSeek | Yes | Same page | No consumer plan |
| Once, inside an FAQ | Yes, one model | Pro only | |
| OpenAI | Page returned 403 | Yes, 1.05M | No |
| xAI | Page returned 403 | Yes, by model | No |
| Perplexity | Page returned 404 | One model only | No |
| Mistral | No figure at all | Third-party models only | No |
| Microsoft | Page returned 404 | No figure | No |
Four of the eight pages a buyer would naturally open returned 403 or 404 to an ordinary request. I am not claiming those pages are broken for humans in browsers. I am saying that the public, checkable record of what these products offer is thinner than the confidence of the articles summarising it.
Why this happened
Context window used to be a clean headline number because a model had one. Model names also churn fast enough that a figure can outlive the model it described, which I tracked separately in nine AI models retired in 2026. Now it varies by model, by product surface, by whether you have enabled usage credits, and sometimes by whether you remember to pick a specific model variant from a slash command.
A single number on a pricing page would now be wrong. So vendors print a ceiling, or a range, or nothing, and push the real answer into documentation that the pricing page does not link to. That is a rational response to a messy reality, and it still leaves the buyer holding a number that is not theirs.
Meanwhile the articles that rank for these queries keep printing one clean figure, because one clean figure is what people search for. Including, until this morning, mine.
What to do before you pay
- Find the per-model table, not the pricing page. The pricing page gives a ceiling. The table gives your number.
- Check which surface that table covers. Chat, agent tool and desktop app can differ by five times on the same model.
- Confirm the specific model by name if that model is the reason you are buying. No checkout will stop you.
- Treat any single number in a third-party article as wrong until a vendor page agrees with it. That includes this one, after enough time has passed.
What I could not verify
Read on 10 November 2026 and refused: openai.com/chatgpt/pricing (403), help.openai.com Plus article (403), x.ai/api (403), microsoft.com plans and pricing (404), perplexity.ai/hub/faq (404), support.anthropic.com Max plan usage article (404 after redirect).
A 403 means a server declined an automated request. A 404 means a URL did not resolve on the day I asked. Neither is evidence about a product, and I have not treated them as any. Where I could not read a page, the row says so rather than borrowing a number from somewhere else.
For OpenAI and Microsoft in particular, that means the consumer context window is simply unreported here. If you need it, open the page in a browser while signed in, which is the only source that reflects your account rather than a marketing page.
Sourcing
Read 10 November 2026, all vendor pages:
- Anthropic: claude.com/pricing, support.claude.com article 8606394 on context window size by model, support.claude.com article 11647753 on usage and length limits, platform.claude.com model overview
- OpenAI: developers.openai.com API model docs
- Google: gemini.google/subscriptions and aistudio.google.com model docs
- xAI: docs.x.ai model and pricing table
- Mistral: mistral.ai/pricing and docs.mistral.ai models overview
- DeepSeek: api-docs.deepseek.com pricing, and deepseek.com
- Perplexity: docs.perplexity.ai model list
- Microsoft: learn.microsoft.com Copilot hub
If you want the same treatment applied to prices rather than specs, I did that in which models cheap AI plans actually run and in the pricing page audit. The method is the same: read the vendor's own page, write down the date, and say so when you could not.
The context window is the one number in AI marketing that sounds like a fact. It turns out to be a ceiling, a product decision, and a footnote, and which of the three you get depends on how far you were willing to click.

