Mistral Large 4's Context Window: 1M Claimed, 524K Measured, and Why the Difference Is Billed

Mistral's model card says one thing and an independent measurement says another, and the gap is not cosmetic. The card lists a 1M-token context window. Artificial Analysis, which evaluates the configuration it can actually run, measured 524,288 tokens. Every token of input you send is metered, so the question "how much context do I have" is also the question "how much am I allowed to spend in one request."

The difference between those two numbers is 2×, which on a model that charges $1.36 per million input tokens is not a rounding error.

What the numbers are, and where each comes from

• 1,048,576 tokens — the value in the config file Mistral ships with the weights. It is a ceiling written by the vendor, and the model card repeats it as "1M".

• 524,288 tokens — what Artificial Analysis measured on the configuration it tested. Half the ceiling.

• What that costs — at $1.36 per million, filling 524K costs about $0.71 per request; filling the full 1M would cost $1.43. Both are the input side of a single call, before any output.

The vendor figure is not a lie and the measured one is not a bug. A context window is a capability of the architecture; what a given serving configuration will actually accept — with its KV-cache budget, its tensor parallelism, its batch size — is smaller, and the gap between the two is normal across this industry. It is just rarely stated on the same line as the price.

Why this matters more than it does elsewhere

Three reasons specific to this model.

It is a trillion-parameter MoE. 1.05 trillion parameters with 52 billion active per token, plus a 1.6-billion-parameter vision encoder. KV cache for a long context on that class of model is the binding constraint on serving, not the config file's intent. The measured number is the honest planning number.

It is priced per token, not per request. A long-context request on this card is a line item that scales linearly with what you send. You are not buying "a 1M window"; you are buying tokens, and the window decides only how many can arrive in one call.

The weights ship with the config. When you download 1,060.6 GB of bf16 and read code>config.json/code>, you will see the 1,048,576 and you will be tempted to plan against it. The vLLM recipe asks for tensor parallelism of 8 with FP8 KV cache, which is exactly the kind of configuration where the served window and the configured window stop agreeing.

How to measure your own ceiling in an afternoon

Send a request that exceeds what you think the limit is and read the error — providers return a token-count message rather than silently truncating in most cases. That gives you the served ceiling on the host you actually use, which is the only number your budget should contain.

Then send a request at 80% of that ceiling and watch the latency. A trillion-parameter model reading a long context has a first-token time that grows with prompt length; the 524K figure is a window, not a promise about speed.

Finally, compare the resulting ceiling against your 95th-percentile input length rather than the average. The average never binds; the tail is where the chunking pipeline you did not budget for appears.

What the competition's windows look like

• DeepSeek V4 Pro — 1M context, 384K output, $0.66 in and $1.98 out off-peak.

• Grok 4.7 — 500K context, 450K output, $2.00 and $6.00 flat.

• Qwen3-Max — 262K context, 65K output, $0.36 and $1.43.

• Mistral Large 4 — 1M claimed, 524K measured, $1.36 and $4.18 at list.

Against that field, Mistral's measured window sits between the 262K and 1M claims — closer to the middle than to the top. If context is the reason you are looking at it, the measured number is the one that belongs in the comparison table, and it costs half the headline.

Mistral Large 4 is not on OrcaRouter's catalogue, so this is not a route we can offer; it is callable through Mistral's own API and third-party hosts. What we do offer generally is one OpenAI-compatible endpoint over 224 models at provider list price with no markup — where the vendor's window, whatever it turns out to be, is the one you get billed against.