context window comparison 2026: most models have a million tokens. few use them well.

here is the short answer to any context window comparison as of september 2026. claude, gpt, gemini and deepseek all ship roughly 1 million tokens of input. grok's newest flagship ships 500k. max output ranges from 64k to 384k tokens.

the numbers have converged, so size is no longer the thing to compare. pricing above 200k tokens differs a lot, and so does how much of that window a model actually uses. the research is blunt on the second point. in chroma's tests, every model did better with a focused prompt than with the full long one. a bigger window does not replace searching your own history first.

how big is each model's context window as of september 2026?

every figure below comes from the vendor's own docs or model card, checked on september 24, 2026. prices are per million tokens, input then output.

sources: anthropic models overview, openai models, gpt-6 astra model page, openai pricing, gemini 3.1 pro model page, gemini 3.8 flash model page, gemini pricing, xai models, xai pricing, deepseek api pricing, deepseek v4 model card, qwen3.8 27b model card, kimi k3 model card, meta's llama 4 post.

three things stand out.

• the token limit is a commodity now. openai's gpt-6 family ships the same 1.05m window on all three tiers. astra launched on september 3, 2026, and sol and luna followed on september 22, per openai's api changelog. deepseek made 1m the default across all its official services on april 24, 2026.
• grok went smaller. xai's newest flagship, grok 4.7, has half the window of the older grok 4.3.
• the 10m number is old. llama 4 scout claimed 10 million tokens on april 5, 2025. every hosted api in this table stops at about 1m.

what does long context pricing actually cost you?

this is where the vendors split, and it matters more than the window size.

anthropic's pricing page says its 1m models are billed at standard rates across the full window: "a 900k-token request is billed at the same per-token rate as a 9k-token request."

openai, google and xai all use a cliff. openai's gpt-6 model pages charge requests over 272k input tokens "2x input and cache rates and 1.5x output for the full request." google re-rates gemini 3.1 pro prompts over 200k from $2 to $4 input. xai doubles grok 4.7 at 200k.

the cliff applies to the whole request, not only the overflow. a 300k prompt to gpt-6 astra costs $6 in input alone, where a 250k prompt costs $2.50. crossing 272k by 28k tokens more than doubles the input bill.

one more wrinkle. anthropic notes that claude 4.7 and later use a tokenizer that "produces approximately 30% more tokens for the same text." so a million claude tokens holds less text than it used to. compare windows in words, not tokens.

how much work fits in a 1 million token context window?

a full window holds far less of your work than the number suggests.

anthropic's pricing page puts an average web page at about 2,500 tokens and a research paper pdf at about 125,000. so a full window holds around 400 web pages, or 8 papers.

a month of what actually crossed your screen does not fit. no window on this list holds it, and at 1m tokens a single stuffed prompt to claude opus 5.5 costs $4 in input every time you ask.

we covered the mechanics in what is a context window.

do models actually use long context well?

the honest answer is partly, and much worse than the spec sheet suggests. four pieces of published evidence, oldest first.

1. lost in the middle, july 6, 2023. the paper by nelson liu and colleagues found models use facts at the start or end of a long input far better than facts in the middle.
2. nolima, february 2025. the nolima benchmark cut the literal word overlap that makes classic needle tests easy. of 13 models claiming 128k or more, "at 32K, for instance, 11 models drop below 50% of their strong short-length baselines." gpt-4o fell from 99.3% to 69.7%.
3. chroma's context rot report, july 14, 2025. chroma tested 18 models and found "model performance varies significantly as input length changes, even on simple tasks." on longmemeval, a focused prompt of about 300 tokens beat the full 113k token conversation.
4. vendor needle scores, 2026. anthropic's opus 4.6 launch post on february 5, 2026 reported 76% on the 8 needle 1m variant of mrcr v2, against 18.5% for sonnet 4.5. deepseek's v4 card reports 83.5 on mrcr 1m for v4 pro in its max mode, in a table where its own run scores opus 4.6 at 92.9. different harnesses, so do not compare the two vendors' numbers directly.

read those vendor numbers carefully. a 76% score on a retrieval test still leaves about a quarter of the available score on the table at full length. and a needle test is the easy version, since the question tells the model what to look for.

anthropic says the quiet part in its own context window docs: "more context isn't automatically better. as token count grows, accuracy and recall degrade, a phenomenon known as context rot." that line is still live on september 24, 2026, next to a 1m token spec.

so which model should you pick for long context?

among the frontier tiers, use claude sonnet 5 when you truly need to paste in a whole corpus. at its flat $2 rate a 900k prompt costs $1.80 in input, against $3.60 on gpt-6 sol or gemini 3.1 pro at their long context rates.

use gpt-6 sol or gemini 3.1 pro if you stay under 200k to 272k tokens, where their base rates are competitive. use gemini 3.8 flash or gpt-6 luna for cheap bulk reading. self host deepseek v4 when the data cannot leave your servers.

and for anything personal, pick retrieval over all of them. in our view the gap that matters is the one chroma measured between a 113k token stuffed prompt and a 300 token focused one, not the gap between models.

that is the whole case for rag and for context engineering. search your own material first, then send the model a handful of passages it cannot miss.

why your context should not live in one model

the table above will be wrong by christmas. openai has shipped two generations since july: gpt-5.6 on july 9, then gpt-6 astra on september 3. anthropic called off a planned september 1 price rise on sonnet 5 before it took effect. models are interchangeable and turn over every quarter. your context should not be locked to one of them.

remynd keeps the record on your side. it is a mac app that captures the focused window, runs ocr locally with apple vision, and keeps the searchable index on your mac. you can exclude apps and sites, and recordings are kept for 30 days by default. call transcription runs on device with mlx.

access to that history is read only. remynd ships an agent window that runs your own claude code or codex cli against it, so you can switch models without losing context. settings also take a custom api endpoint compatible with the openai responses api.

to be precise: the archive lives on your mac, but that does not mean nothing leaves it. whatever you send to a hosted model goes to that provider. our security page and private ai on your mac spell out the details.

download remynd for mac and let your own agent search your week instead of guessing at it.