What a Million-Token Window Actually Holds
One million tokens is now the common figure on flagship model pages. It is not a common measurement.
This note reads the published token limits for the current flagship models at three vendors, as the documentation showed them on October 27, 2026.

The three windows are close in size and different in kind. One splits input from output, one caps input below the window, and one shares a single budget across everything.
Anthropic publishes a window and a separate output cap
The Anthropic models overview lists four current models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5.

All four show the same two rows. Context window is 1M tokens. Max output is 128K tokens.
The output cap is additional to the window rather than carved out of it, which is the simplest of the three arrangements to budget for.
All four models carry identical token limits, which is itself a finding. Within this family, choosing a model is a decision about price and latency, not about how much it can read.
The page prints the price difference beside the limits. Input ranges from $0.10 per million tokens on Haiku 5.5 to $10 on Fable 5.1, with the same window on both.
The page also points to a programmatic route. The Models API returns max_input_tokens, max_tokens and a capabilities object for every available model.
That is worth knowing because a documentation table is a snapshot and an API response is not. A team shipping against these limits can read them at runtime.
OpenAI publishes a window larger than its own input limit
The GPT-6 Astra model page lists three separate figures where most vendors list two.
- 1,050,000 context window
- Maximum input tokens: 922,000
- 128,000 max output tokens
Those three numbers do not add up to each other, and they are not meant to. The window is the total, the input ceiling is lower, and output is capped separately.
GPT-6 Luna publishes the identical set: 1,050,000 window, 922,000 maximum input, 128,000 maximum output. Only the knowledge cutoff differs, April 30, 2026 for Astra and May 18, 2026 for Luna.
The practical reading is that 922,000 is the number a document pipeline has to respect. The 1,050,000 figure is the headline.
It is also the figure to compare against the price threshold. Both models reprice above 272,000 input tokens, as set out in the long-context price note.
Gemini shares one budget across input, thinking and output
The Gemini 4 Argon model page labels its token limit unusually, and the label is the whole story.
Context window (unified input and output token limit): 1,048,576. Max output: 65,536 by default for synchronous calls, or up to 1,048,576 with long decode.
Unified means the input and the output draw on the same allowance. A 900,000-token input leaves roughly 148,000 tokens of room for everything the model produces.
That subtraction is editorial, taken from the published window of 1,048,576. The page gives the total and leaves the remainder to the reader.
The arrangement also changes how a long conversation behaves. Each turn adds to the same budget, so a thread that keeps its full history shrinks its own room to answer.
The long decode documentation makes the sharing explicit. It describes 1,048,576 tokens shared across input, thinking and output.
Thinking tokens sitting inside the same budget is the detail most likely to surprise. A reasoning-heavy answer consumes window as well as money.
Long decode is the same window read as output
The default synchronous output cap on Gemini 4 Argon is 65,536 tokens, which is half of what the other two vendors allow in one call.
Long decode lifts it by stitching calls together. A single model call generates up to 64K tokens, and the server continues from where the last one stopped.
To use it, the documentation says to create an interaction with background set to true and max_output_tokens greater than 65,536, up to 1M.
Each hop adds a model_output step to interaction.steps under the same interaction ID, so the caller joins those steps to read the full response.
So the million-token output figure is real, and it is not a single generation. It is a server-side loop that a buyer should cost and time as a loop.
A million tokens is not a million tokens of text
One line on the Anthropic pricing page changes how every window figure above should be read.
Claude 4.7 and later models, and Claude Mythos Preview, use a newer tokenizer. The page states it produces approximately 30 percent more tokens for the same text.
Claude Sonnet 4.6 and earlier models use the previous tokenizer, per the same page. The exact increase depends on the content and workload shape.
Read as capacity, a 1M window on the newer tokenizer holds noticeably less text than a 1M window on the older one. That is editorial arithmetic from the published 30 percent figure.
Read as cost, the same document bills more tokens at the same per-token rate. The rate card did not move and the invoice does.
This is why comparing two windows by their headline number alone is unsafe. Tokens are a vendor-specific unit, not a standard one.
It also makes a year-on-year price comparison harder than it looks. A rate card that holds steady while the tokenizer changes is a quiet increase on the same work.
Anthropic says the newer tokenizer contributes to improved performance across a wide range of tasks. The point here is not that the trade is bad, only that the unit moved.
How to compare two windows honestly
- Ask whether the window includes output, because on one of these three platforms it does
- Find the maximum input figure separately, since OpenAI publishes 922,000 against a 1,050,000 window
- Check the default output cap before the maximum one, because 65,536 is what a plain synchronous call gets on Gemini 4 Argon
- Treat thinking tokens as window and as spend, since at least one vendor counts them in both
- Test with your own documents rather than a token estimate, because tokenizers differ by about 30 percent on one platform's own admission
- Read the limits from the API where the vendor offers it, so the figure matches the model you are actually calling
Bottom line
The three flagship windows read 1,000,000, 1,050,000 and 1,048,576 tokens. The differences in the headline are trivial and the differences underneath are not.
Claude gives a separate 128K output allowance. GPT-6 caps input at 922,000. Gemini 4 Argon shares one budget and defaults output to 65,536.
For what this capacity costs once a prompt gets long, see the long-context price note and the cached-input price note.
For the monthly ceiling the spend runs into, see the monthly cap note. For how long each model stays callable, see the model shutdown calendar.
Coding tools that lean hardest on long context are graded in the Cursor review and the GitHub Copilot review, alongside the shortlist of AI assistant plans.
What this brief prints and what it cuts is in issue zero.
Sources
- Anthropic models overview (context window and max output) - checked October 27, 2026
- Anthropic model pricing (tokenizer note) - checked October 27, 2026
- OpenAI GPT-6 Astra model page - checked October 27, 2026
- Gemini 4 Argon model page - checked October 27, 2026
- Gemini API long decode documentation - checked October 27, 2026