Mistral Large 4 vs Qwen3-Max: The Cheap Model Is the One With the Small Window

The price gap runs the opposite way to the size gap, which is the whole interest of this pairing. Qwen3-Max costs $0.36 per million input tokens and $1.43 output — a quarter of Mistral Large 4's list output price and about a third of its promotional one. It serves a 262K-token context with a 65K output ceiling. Mistral Large 4 claims a 1M context, measures 524,288 on the tested configuration, and produces up to that much in a single answer.

So the cheap model is the small-window model, and whether that matters depends entirely on what you send it.

The cards

• Qwen3-Max — $0.36 in, $1.43 out per million; 262K context; 65K maximum output.

• Mistral Large 4 — $1.36 in, $0.14 cached, $4.18 out; promoted at $0.68 / $0.07 / $2.09. 1M context claimed, 524,288 measured; 1.05 trillion parameters, 52 billion active, with a 1.6-billion-parameter vision encoder.

• The ratio — Mistral is 3.8× Qwen on input and 2.9× on output at list; 1.9× and 1.5× on promotion.

Those two ratios are the same product described two ways, and only one of them survives the end of a promotion nobody has dated.

Where 262K actually binds

Two hundred and sixty-two thousand tokens is roughly 200,000 words. A medium codebase, a long contract set, a multi-hour transcript — all of those cross it. A single document, a chat history, an agent's working set — most of those do not. The honest way to decide is to measure the 95th percentile of your input length and compare it against the ceiling rather than against the average; the average never binds and the tail always does.

The output ceiling binds far more often than people expect. Sixty-five thousand tokens is about 50,000 words, and any single-request generation of a long report, a large code file, or a full transcript synthesis approaches it. Mistral's 524K measured window is eight times larger, which is the difference between one call and a chunking pipeline — and a chunking pipeline costs engineering time that no per-token price reflects.

What the price difference actually pays for

Run one million input tokens and 100,000 output tokens through both cards. Qwen3-Max: $0.36 plus $0.143 — about 50 cents. Mistral at list: $1.36 plus $0.418 — $1.78. On promotion: 89 cents.

So the flagship costs between 1.8× and 3.6× the cheaper model per identical task, and it buys a context nearly four times larger, an output ceiling eight times larger, native image input via a 1.6-billion-parameter encoder, published weights and a licence you can read. Whether that is worth $1.28 on this task is a question about the task, not about the market.

One thing the comparison does not capture: Mistral publishes a $0.14 cached-input line, and Qwen3-Max's card as we read it does not carry an equivalent. For an agent that re-sends a stable prefix on every turn, that missing row can be worth more than the sticker difference — and it is the first thing to measure rather than assume when testing both.

Which one to route where

Qwen3-Max is callable on OrcaRouter at its published rate, on the same OpenAI-compatible endpoint as a 224-model catalogue, with automatic failover. Mistral Large 4 is not in that catalogue; it runs through Mistral's own API and third-party hosts.

The routing rule the cards support is simple: short-input, short-output, high-volume work goes to the 36-cent model, because none of the ceilings bind and the price is the only variable. Anything in the long tail — a full document, a large codebase, a single-call generation — goes to the model with the larger window, because the cheaper one will refuse the job rather than discount it.

And if you are choosing Mistral on the promotional number, remember what this page compared: at list, the ratio against Qwen is nearly four to one on input, not two.