No one sells a real 10M-token window — and eleven of thirty-seven models charge more before you reach the max
August 28, 2026 · Updated October 8, 2026
"Models with 10 million token context window" is tied for the highest-volume single spec query in this site's keyword research. Here is the answer nobody selling one will give you: of the 37 models tracked here, none ships a 10M-token window — native, extended, or otherwise. The biggest real number on sale is 1,050,000 tokens, and across the 37 tracked models the spec sheet splits three ways: 25 models charge one flat rate all the way to their maximum, 11 quietly move you to a higher price once your prompt crosses a threshold the marketing page doesn't lead with, and 1 ships a smaller native window that only reaches 1M if you re-configure it yourself. Every number below was checked against the vendor's own documentation on 2026-08-28 (Claude Fable 5.1 added 2026-09-02, Gemini 3.8 Flash and Muse Spark 1.3 both added 2026-09-03, GPT-6 Astra added 2026-09-04, DeepSeek V4.1 Flash added 2026-09-10, Step 5 Preview and both MiMo-V2.6 models added 2026-09-22, Grok 4.7, Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna added 2026-09-23, Tencent Hy4 preview added 2026-09-27, Claude Sonnet 5.5 added 2026-09-29, GPT-6.1 Sol added 2026-09-30, Gemini 4 Argon added 2026-10-02, Mistral Large 4 added 2026-10-06, Claude Haiku 5.5 added 2026-10-08), and each table names the vendor pages it came from.
The largest context windows on sale, by model
The largest context window on sale from any model tracked here is 1,050,000 tokens, and no model reaches 10M natively or by extension. Every tracked model, grouped by the window its vendor states — with one exception noted in the table below, since Gemini 4 Argon has been announced but has no API to buy and no vendor-stated window and by whether that window is priced at one rate — the two detailed tables further down carry each row's threshold, rates and source:
| Window (tokens) | Models | Priced how |
|---|---|---|
| 1,050,000 | GPT-5.6 Luna, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Luna, GPT-6 Sol, GPT-6.1 Sol | tiered above 272K input |
| 1,048,576 | Gemini 3.1 Pro Preview | tiered above 200K prompt |
| 1,000,000 | Claude Haiku 5.5 | tiered above 100K prompt |
| 1,048,576 | GLM-5.3-Flash, Gemini 3.7 Flash, Gemini 3.8 Flash, Kimi K3, MiMo-V2.6-Flash, MiMo-V2.6-Pro, Muse Spark 1.2, Muse Spark 1.3, Tencent Hy4 preview | flat |
| 1,000,000 (Argon: not vendor-stated, not on sale) | Gemini 4 Argon, MiniMax M3 | Argon: announced flat rate, $2/$10 introductory then $4/$20 · MiniMax: tiered above 512K input |
| 1,000,000 | Claude Fable 5, Claude Fable 5.1, Claude Opus 4.8, Claude Opus 5, Claude Opus 5.5, Claude Sonnet 5, Claude Sonnet 5.5, DeepSeek V4 Flash (0731), DeepSeek V4 Pro (0813), DeepSeek V4.1 Flash, GLM-5.2, GLM-5.3, Mistral Large 4, Qwen3.8-Max, Step 5 Preview | flat |
| 500,000 | Grok 4.6, Grok 4.7 | tiered from 200K prompt, inclusive |
| 262,144 native | Qwen3.8-Flash-Next | no first-party API; 1M only via YaRN if you serve it yourself |
Where the 10M number actually comes from
A context-window figure can mean two different things, and the gap between them is where the biggest claims live. A native window is what the model was trained to handle and what the default configuration serves. An extended window is what you can reach by stretching the model's position encoding — techniques like RoPE scaling and YaRN — which the vendor may support, recommend, or merely permit, and which the model was not trained at.
The cleanest worked example on this site is Qwen3.8-Flash-Next. Its Hugging Face card states a native length of 262,144 tokens, "extensible up to 1,000,000 tokens" via YaRN — the vendor's own words, and an honest framing. The same card describes a separate sibling product — the bare Qwen3.8-Flash — with "1M context length by default," and that sibling went on sale through Alibaba's cloud API on 2026-08-26 (OpenRouter lists the live maker-operated endpoint at a 1,000,000-token window; we found no first-party Alibaba announcement page). Those are three different claims — 262K trained on this checkpoint, 1M reachable on it with work, and 1M standard on a different product — and a spec-table that prints "1M" for this model collapses all three into one number.
That is the mechanism behind every headline-grabbing 10M figure we could find: a length-extension claim, not a trained length. No model tracked here makes even the extension claim at 10M, so this table stops at what is actually on sale.
The eleven models where 1M costs more than the price page led with
Ten tracked models publish a per-token rate, and then a second, higher rate that kicks in at or above a prompt-length threshold. The step isn't a secret — the thresholds and both rates below are from each vendor's own pricing page — but every one of them quotes the lower rate first. If your workload lives in long context, the second column is your real price.
| Model | Window | Threshold | Below → above (per 1M tokens) | Source, checked 2026-08-28 |
|---|---|---|---|---|
| GPT-5.6 Sol | 1,050,000 | >272K input | $4.00/$20.00 → $8.00/$30.00 (2× in, 1.5× out) | developers.openai.com |
| GPT-5.6 Luna | 1,050,000 | >272K input | $0.20/$1.20 → $0.40/$1.80 (same family rule) | developers.openai.com |
| GPT-6 Astra | 1,050,000 | >272K input | $10.00/$50.00 → $20.00/$75.00 (same family rule) | developers.openai.com, checked 2026-09-04 |
| GPT-6 Sol | 1,050,000 | >272K input | $2.00/$10.00 → $4.00/$15.00 (same family rule) | developers.openai.com, checked 2026-09-23 |
| GPT-6 Luna | 1,050,000 | >272K input | $0.10/$0.50 → $0.20/$0.75 (same family rule) | developers.openai.com, checked 2026-09-23 |
| GPT-6.1 Sol | 1,050,000 | >272K input | $2.00/$10.00 → $4.00/$15.00 (same family rule) | developers.openai.com, checked 2026-09-30 |
| Gemini 3.1 Pro Preview | 1,048,576 | >200K prompt | $2.00/$12.00 → $4.00/$18.00 | ai.google.dev pricing |
| Grok 4.6 | 500,000 | ≥200K prompt | $2.00/$6.00 → $4.00/$12.00 (2× everything) | docs.x.ai |
| Grok 4.7 | 500,000 | ≥200K prompt | $2.00/$6.00 → $4.00/$12.00 (2× everything) | docs.x.ai, checked 2026-09-23 |
| MiniMax M3 | 1,000,000 | >512K input | $0.30/$1.20 → $0.60/$2.40 (all three rates step up) | minimax.io model page + pricing guide |
| Claude Haiku 5.5 | 1,000,000 | >100K prompt | $0.10/$0.50 → $0.50/$2.50 (5× both, the lowest threshold on this list) | platform.claude.com model overview, checked 2026-10-08 |
Three details worth reading twice. Grok 4.6's boundary is inclusive — xAI's own row label reads "≥ 200k", so a prompt of exactly 200,000 tokens already bills at the doubled rate (Grok 4.7 carries the identical rule and threshold). MiniMax's table steps up all three of its rates — input, output, and cache reads — for requests above 512K input; unlike xAI and OpenAI, its page never states whether tokens below the line keep the lower rate. And OpenAI's rule applies "for the full request" too: crossing 272K doesn't split the bill, it doubles the input side of all of it — the same rule GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol carry, confirmed on their own pricing pages, and none is a promotional rate: OpenAI's own comparison table frames GPT-6 Sol and Luna's prices as the new standard, calling their GPT-5.6 predecessors' rates the promotional ones instead, and GPT-6.1 Sol's announcement calls its $2/$10 its "standard API prices."
One more pricing footnote: three of the eleven base rates above are themselves promotional — OpenAI labels GPT-5.6 Sol's $4.00/$20.00 a promo guaranteed at least through 2026-11-21, calls GPT-5.6 Luna's $0.20/$1.20 promotional too (its own comparison table for GPT-6 Sol and Luna cites both prior rates that way), and MiniMax badges its $0.30/$1.20 as a "Permanent 50% off" rate against list prices of $0.60/$2.40 and $1.20/$4.80 per tier.
None of this makes the windows fake. It makes the phrase "1M context window at $2 per million tokens" incomplete — the window is real, the price applies to less of it than the sentence implies.
The twenty-five that charge one rate to the ceiling
The other 25 models with published pricing charge the same per-token rate at token 1,000,000 as at token 1,000 — confirmed against each vendor's own pricing page, not assumed from the absence of a disclosed tier, with one exception: Gemini 4 Argon has no pricing page and no API to buy, so its row rests on Google's announcement disclosing no length threshold, and its window figure is Artificial Analysis's rather than Google's.
| Model | Window (vendor-stated) | Flat-rate note, checked 2026-08-28 |
|---|---|---|
| Claude Fable 5 | 1,000,000 | Anthropic bills a 900K-token request at the same per-token rate as a 9K one (its own docs' example, Claude 4.6+); speed and batch variants exist, none keyed to length |
| Claude Fable 5.1 | 1,000,000 | Same rule — carries the price and the pricing structure over unchanged from Claude Fable 5, per Anthropic's own models comparison table, checked 2026-09-02 |
| Claude Opus 5 | 1,000,000 | Same rule |
| Claude Opus 5.5 | 1,000,000 | Same rule — Anthropic's pricing page states the full window bills at one rate regardless of length, checked 2026-09-23 |
| Claude Opus 4.8 | 1,000,000 | Same rule |
| Claude Sonnet 5 | 1,000,000 | Same rule |
| Claude Sonnet 5.5 | 1,000,000 | Same rule — Anthropic's "What's new in Claude Sonnet 5.5" page states the same prices as Sonnet 5 including caching and batch, and the pricing page bills the full window at one rate for every model from Claude 4.6 on, checked 2026-09-29 |
| DeepSeek V4 Pro (0813) | 1,000,000 | Rates vary by time of day, never by prompt length; output capped at 384K |
| DeepSeek V4 Flash (0731) | 1,000,000 | Same peak/off-peak structure, no length tier |
| DeepSeek V4.1 Flash | 1,000,000 | Same peak/off-peak structure as the earlier V4 models, no length tier, checked 2026-09-10 |
| Step 5 Preview | 1,000,000 | One row on StepFun's USD price list; no length, time-of-day or promotional condition anywhere in its docs, checked 2026-09-21. Maximum output 64K tokens |
| Kimi K3 | 1,048,576 | Single rate on Moonshot's pricing page |
| GLM-5.3 | 1,000,000 ("1M") | Single rate on docs.z.ai |
| GLM-5.3-Flash | 1,048,576 ("1M") | Single rate; launch promo aside, no length tier |
| GLM-5.2 | 1,000,000 ("1M") | Single rate |
| Qwen3.8-Max | 1,000,000 | One rate per region — Singapore bills above the other five — but no length tier |
| Muse Spark 1.2 | 1,048,576 ("1M") | No length tier on Meta's pricing table |
| Gemini 4 Argon | 1,000,000 (not vendor-stated) | No length tier disclosed; the only rate change is introductory — $2/$10 until an unnamed intro period expires, then $4/$20. Google's post gives a 1M output ceiling, not a context window; the window here is Artificial Analysis's figure, checked 2026-10-02 |
| Muse Spark 1.3 | 1,048,576 ("1M") | Same rule as Muse Spark 1.2 — Standard vs. Contributor is a training-data-use split, not a length tier, checked 2026-09-03 |
| Gemini 3.7 Flash | up to 1,048,576 | The explicit exception in Google's own lineup: no length-based price split, unlike its Pro sibling |
| Gemini 3.8 Flash | up to 1,048,576 | Same rule as Gemini 3.7 Flash — no length-based tier on Google's pricing page for this model, checked 2026-09-03 |
| Mistral Large 4 | 1,000,000 | No length tier disclosed; the docs pricing page lists $0.68/$2.09 per 1M tokens (cached $0.07) with no long-context condition. Artificial Analysis separately lists a 524k window for this model — the 1M here is the vendor's stated spec, checked 2026-10-06 |
| MiMo-V2.6-Pro | 1,048,576 | Single rate on Xiaomi's model page, no length threshold or time-of-day condition, checked 2026-09-22 |
| MiMo-V2.6-Flash | 1,048,576 | Single rate on Xiaomi's model page, no length threshold or time-of-day condition, checked 2026-09-22 |
| Tencent Hy4 preview | 1,048,576 ("1M") | One rate on Tencent Cloud's TokenHub price list, no length or peak/off-peak condition, checked 2026-09-27. Maximum output 64K tokens |
Two attribution notes, because precision is the point of this page: several vendors' pages say only "1M" — where they do, the table marks it ("1M") and the integer shown follows that model's own config or API reference, which is why the same label renders as 1,000,000 on some rows and 1,048,576 on others. And Gemini 3.7 Flash's card lists 1M with no length-based price split — Google's own stated structure for that model, sitting one product away from the Pro model that does carry a 200K split.
Qwen3.8-Flash-Next sits outside both tables: it has no first-party API to price at all — 262,144 tokens native as open weights, 1M reachable via YaRN if you serve it yourself.
How to read a context-window claim now
Three questions turn any context-window spec into something you can actually budget against. Is it native or extended? — if the card says "extensible," the headline number was not trained, and the two claims are recorded separately here. Is the rate flat across the window? — eleven of thirty-seven models here say no once you find the right table row, and two of them start charging double at exactly the threshold, not past it. Who says so, and when? — every claim above carries the vendor page it came from and the date we read it, which is also why this piece publishes no blended "context quality score": a composite number would hide exactly the distinctions this table exists to show. The per-model records — pricing, scores, and what changed since — update on each model's own page before they appear anywhere else; start from the full catalog.
Context-window questions, answered
Do any AI models have a 10 million token context window?
No. Of the 37 models tracked here, none ships a 10M-token window — native, extended, or otherwise. The biggest real number on sale is 1,050,000 tokens; the rest of the 10M story is marketing arithmetic on output limits and extensible configs, covered above.
Which model has the largest context window on sale?
The six GPT-family rows — GPT-5.6 Sol and Luna, GPT-6 Astra, GPT-6 Sol and Luna, and GPT-6.1 Sol — all list 1,050,000 tokens. That is the ceiling of every price page this piece checked; nothing currently listed goes higher.
Which models charge more once a prompt gets long?
Eleven of the 37. The GPT-5.6/6 Sol–Luna line carries an identical >272,000-input threshold, Grok 4.6 and 4.7 switch to a higher tier at ≥200,000 prompt tokens, and Claude Haiku 5.5 steps up above 100,000. The full eleven, with each threshold and its source, are in the tiered table above.
What happens at the 272K input threshold?
The per-token rate steps up once input crosses a threshold. Where the line sits, and whether the higher rate begins at it or past it, differs by family: the six GPT-family rows step up past 272,000 input tokens, while Grok 4.6 and 4.7 switch at 200,000 — and those two begin charging the higher rate at exactly the threshold, not past it. That boundary wording is one of the three reading rules the closing section walks through.
What is the difference between a native and an extensible window?
If a model card says the window is "extensible," the headline number was not trained — the base model ships smaller and re-configuration reaches the rest. One of the 37 tracked here, Qwen3.8-Flash-Next, ships a 262,144-token native window that only reaches 1M if you re-configure it yourself.
Do Grok models have a long-context price tier?
Yes. Grok 4.6 and Grok 4.7 both carry a ≥200,000-prompt threshold on xAI's own model table — past it, the per-token rate steps up. The window itself holds at 500,000 tokens on both.
How many models charge one flat rate all the way to the ceiling?
Twenty-five of the 37 — the full list with per-model rates is the second table above. Their shared property is that the number on the pricing page is the number you pay at maximum context, with no threshold row to find.
Was every context figure on this page stated by the vendor?
All but one. Each figure was read from the vendor's own documentation on the dated checks in the sources note below — except Gemini 4 Argon's window, where Google publishes a 1M output-token limit rather than a context window, so that row follows Artificial Analysis's measurement and says so in the table.
Revision log
- Updated again October 8, 2026: Claude Haiku 5.5 added as the 37th tracked model, one day after launch. Anthropic's model overview prices it at $0.10/$0.50 per 1M tokens for prompts up to 100,000 and $0.50/$2.50 above that — a 5× step at the lowest threshold on either table — the biggest proportional step of the eleven, and the lowest entry point. It therefore joins the tiered table, moving that bucket 10 to 11 of 37; the flat-rate (25) and native-smaller (1) buckets are unchanged. Anthropic's own note that about 90% of requests to its previous Haiku fell under 100K tokens is recorded on the model's page, since it is a claim about traffic mix rather than about this table.
- Updated again October 6, 2026: Mistral Large 4 added as the 36th tracked model, on its launch day — it entered public preview the same day. Its docs state the context as 1M tokens and the pricing page lists $0.68/$2.09 per 1M tokens with no length condition, so it joins the flat-rate table, moving that bucket 24 to 25 of 36; the tiered (10) and native-smaller (1) buckets are unchanged. Unlike Gemini 4 Argon, the window here is vendor-stated, so no third-party-figure caveat applies. Mistral also discloses priority-tier rates of $1.36/$4.18 — the figures Artificial Analysis bills — recorded on the model's own page, not in this piece's tables.
- Updated September 2 and twice on September 3, 2026: Claude Fable 5.1, then Gemini 3.8 Flash, then Muse Spark 1.3 shipped as this site's 20th, 21st, and 22nd tracked models. None of the three pricing pages adds a length-based threshold — Fable 5.1 carries Claude Fable 5's flat rate unchanged, Gemini 3.8 Flash carries Gemini 3.7 Flash's, and Muse Spark 1.3's Standard/Contributor split (unchanged from 1.2) is a training-data-use choice, not a length tier — so the flat-rate count in the body above moves from 13 to 14 to 15 to 16 of 22; the five price-tiered models and the one native-smaller model are unchanged.
- Updated again September 4, 2026: GPT-6 Astra shipped as the 23rd. Unlike the three prior additions, its own pricing page discloses the identical >272K-input threshold GPT-5.6 Sol and Luna already carry here — so it joins the tiered table, not the flat-rate one, moving that bucket from 5 to 6 of 23.
- Updated again September 10, 2026: DeepSeek V4.1 Flash shipped as the 24th, replacing DeepSeek V4 Flash per DeepSeek's own changelog.
- Corrected September 17, 2026: the same changelog states V4 Pro API service continues past September 14 with billing unchanged; the earlier claim on this page that V4.1 Flash replaced both models is withdrawn. V4.1 Flash's official USD pricing page carries the same peak/off-peak-by-time-of-day scheme as the earlier V4 models, no length-based threshold — so the flat-rate count moves from 16 to 17 of 24; the tiered (6) and native-smaller (1) buckets are unchanged.
- Updated again September 22, 2026: StepFun's Step 5 Preview shipped as the 25th. Its own USD price list has a single row for it, with no length threshold, time-of-day rate or promotional price on any page of StepFun's docs, so the flat-rate count moves from 17 to 18 of 25. Later the same day Xiaomi shipped MiMo-V2.6-Pro and MiMo-V2.6-Flash as the 26th and 27th tracked models; both official model pages show one rate with no length threshold or time-of-day condition, so the flat-rate count moves from 18 to 20 of 27 — the tiered (6) and native-smaller (1) buckets are unchanged.
- Updated again September 23, 2026: Grok 4.7 shipped as the 28th tracked model. Its own pricing page (docs.x.ai/docs/models) carries the identical >=200K-prompt threshold Grok 4.6 already has here — so it joins the tiered table, not the flat-rate one, moving that bucket from 6 to 7 of 28; the flat-rate (20) and native-smaller (1) buckets are unchanged.
- Updated again the same day (second batch): Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna shipped as the 29th, 30th, and 31st. Opus 5.5's own pricing page states its full 1M-token window bills at one flat rate — into the flat-rate table, moving that bucket 20 to 21. GPT-6 Sol and GPT-6 Luna both carry the identical >272K-input threshold their GPT-5.6 and GPT-6 Astra siblings already have here — into the tiered table, moving that bucket 7 to 9 of 31; the native-smaller (1) bucket is unchanged.
- Updated again September 27, 2026: Tencent Hy4 preview added as the 32nd. Tencent Cloud's TokenHub price list shows one rate for it with no length or peak/off-peak condition, so it joins the flat-rate table, moving that bucket 21 to 22 of 32; the tiered (9) and native-smaller (1) buckets are unchanged.
- Updated again September 29, 2026: a summary table of every tracked model, grouped by window size and pricing rule, now opens the piece; the two detailed tables and every count in the body above are unchanged.
- Updated again September 29, 2026 (later the same day): Claude Sonnet 5.5 shipped as the 33rd tracked model at Sonnet 5's unchanged $2/$10 with the full 1M window at one rate, so it joins the flat-rate table, moving that bucket 22 to 23 of 33; the tiered (9) and native-smaller (1) buckets are unchanged.
- Updated again September 30, 2026: GPT-6.1 Sol shipped as the 34th tracked model. Its own model page carries the identical >272K-input threshold every GPT-6 and GPT-5.6 model here already has, so it joins the tiered table, moving that bucket 9 to 10 of 34; the flat-rate (23) and native-smaller (1) buckets are unchanged.
- Updated again October 2, 2026: Gemini 4 Argon added as the 35th tracked model after Google announced it on September 30. Its announcement discloses no length-based threshold and the only rate change it describes is introductory, with no end date given and no time-of-day structure, so it joins the flat-rate table, moving that bucket 23 to 24 of 35; the tiered (10) and native-smaller (1) buckets are unchanged. It is also the first row here whose window is not vendor-stated: Google gives a 1M output token limit, not a context window, so the 1,000,000 in that row is Artificial Analysis's figure, and the row says so.
Sources: every context-window figure, threshold, and rate above is from the named vendor's own documentation — Anthropic's model comparison table, developers.openai.com model pages, ai.google.dev's pricing page and models reference, the Gemini 3.7 and 3.8 Flash model cards, Google's Gemini 4 Argon announcement for its published rates, and Artificial Analysis's Argon model page for that model's context figure, docs.x.ai's model table, MiniMax's own M3 model page and platform.minimax.io's pricing guide, the Gemini API models table (source of the 3.1 Pro window figure), api-docs.deepseek.com's pricing page, Moonshot's chat-k3 docs, docs.z.ai's pricing pages, Alibaba Cloud Model Studio's qwen3-8-max page, developer.meta.com's Muse Spark page, and the Qwen/Qwen3.8-Flash-Next Hugging Face card — each read on 2026-08-28 (Claude Fable 5.1's row added 2026-09-02; Gemini 3.8 Flash's and Muse Spark 1.3's rows both added 2026-09-03; GPT-6 Astra's row added 2026-09-04, developers.openai.com/api/docs/pricing; DeepSeek V4.1 Flash's row added 2026-09-10, api-docs.deepseek.com/quick_start/pricing; Step 5 Preview's row added 2026-09-22, platform.stepfun.ai pricing and model pages, read 2026-09-21; Grok 4.7's row added 2026-09-23, docs.x.ai/docs/models; Claude Opus 5.5's, GPT-6 Sol's, and GPT-6 Luna's rows all added 2026-09-23, platform.claude.com/docs/en/about-claude/pricing and developers.openai.com/api/docs/models; Tencent Hy4 preview's row added 2026-09-27, Tencent Cloud TokenHub pricing; GPT-6.1 Sol's row added 2026-09-30, Gemini 4 Argon's row added 2026-10-02, developers.openai.com/api/docs/models/gpt-6.1-sol, with the "standard API prices" wording from openai.com/index/introducing-gpt-6-1-sol). Per-model provenance for benchmark scores lives on the linked model pages.