Mistral Large 4 vs Kimi K3.1 Pro: Two Multimodal Flagships, One Licence Apart

Both of these models accept an image and both are sold as flagships by labs outside the American big four. Almost everything else diverges. Mistral Large 4 lists at $1.36 per million input and $4.18 output — currently half that, with no published end date — under a licence that reserves model-as-a-service for a separate agreement. Kimi K3.1 Pro costs $2.00 and $10.00 flat, with weights promised by 4 November and, as of today, an Artificial Analysis label that still reads proprietary. One is Western open weights with a commercial clause; the other is an API-only flagship about to become open weights next week.

The cards

• Mistral Large 4 — 1.05 trillion parameters with 52 billion active by Mistral's card (49 billion by Artificial Analysis), a 1.6-billion-parameter vision encoder, 1M context claimed with 524,288 measured on the tested configuration. $1.36 / $0.14 cached / $4.18 per million, promoted at $0.68 / $0.07 / $2.09.

• Kimi K3.1 Pro — 1M context, 943K maximum output, text, image and video in. $2.00 in and $10.00 out per million on OrcaRouter, at Moonshot's list price. Measured p50 first token: 7.27 seconds.

• The licensing clock — Mistral's weights are live now, 1,060.6 GB in bf16, ungated, under the Mistral Ecosystem Open-Weight License. Kimi's are due by 4 November; whether the licence is MIT-shaped or restricted is not yet public.

Price per token is the wrong unit here

Compare the output lines and Mistral looks 58% cheaper at list and 79% cheaper on promotion. Compare them per completed task and the ranking can invert, because the two models do not spend the same number of tokens getting to an answer.

Mistral's own reporting puts Large 4 at 47% fewer tokens per task than a comparable competitor on Artificial Analysis's coding agent index. If Kimi spends twice the output tokens per task, a $2.09 output line on promotion costs more per finished job than a $10.00 line that finishes in half the tokens. That is not a prediction about which model is better — it is a statement that a per-token comparison without a per-task measurement is a comparison of stickers.

What the cards do tell you honestly is the ceiling. Kimi's 943K maximum output is more than double Mistral's measured 524K context, and a single-call document generation that needs 200,000 output tokens is possible on both while a 600,000-token one is possible on one. If your workload is long-form in a single request, that column decides and the price column does not.

What the licence actually changes for a buyer

Mistral's clause 2 is the reason this comparison is not just arithmetic. Anyone operating a "Model as a Service" business — giving third parties access to the model for inference or post-training in a way that gives them meaningful control over inputs, parameters or training data — must sign a separate commercial licence first. Three carve-outs matter: a product whose capability is embedded in specific features is fine, mere relaying to models hosted elsewhere is fine, and internal use is fine.

Kimi K3.1 Pro's licence is simply not knowable yet. Moonshot has said the weights arrive by 4 November; it has not said under what terms, and the earlier K3 release is no guide, because a licence is read off the release in front of you. Anyone planning to fine-tune and resell should treat this as an open item with a date on it.

Which one to put behind traffic this week

If you need image input and a European-trained model you can inspect, Mistral Large 4 is callable today through Mistral's API and third-party hosts — not through us; it is not in OrcaRouter's catalogue. If you want a model on one key alongside everything else you run, Kimi K3.1 Pro is on OrcaRouter at list price, with automatic failover and a 224-model catalogue behind the same endpoint.

If you are undecided, the useful experiment is the same for both: measure output tokens per completed task on your own workload before you believe either price. Mistral's 47% claim and Kimi's token-heavy long-output ceiling point in opposite directions, and only one of them will be true for your prompts.

What to watch next: whether Kimi's 4 November weights land under a permissive licence, which would remove the only structural advantage Mistral has on this page, and whether Mistral's half-price promotion ends, which would double the gap in the other direction overnight.