Mistral Large 4 goes open-weight, with cached tokens at a tenth of the price

Mistral put the open weights of Mistral Large 4 on Hugging Face on 29 October 2026, after a public preview earlier in the month. “Open-weight” means anyone can download the trained model files. According to Mistral’s model card, it is a multimodal mixture-of-experts model: it works with more than text, and each token goes through only some of the model’s parts. It has about 1 trillion parameters in total, 49 billion of them active for each token (52 billion if embeddings and output layers are counted, which is where the A52B in the repo name comes from), and a context window of 1 million tokens. A token is a small chunk of text, roughly a word fragment. For anyone who pays per token, the price list matters as much as the size.

What it means

Only 49 billion of the parameters run for each token, which probably helps keep the cost of serving it down. None of our sources says why Mistral chose these prices. Mistral lists $0.70 per million input tokens, $2.10 per million output tokens and $0.07 per million cached input tokens. Artificial Analysis, an independent evaluator, lists the same three numbers.

Two things stand out. Cached input costs one tenth of normal input, a 90% discount on cached prompt tokens. And output is the most expensive line, at three times the input price.

Here is why that matters. An AI agent often sends the same long prompt, with instructions, files and history, again and again, and gets a short reply each time. That is the case where caching should pay off. The chart below is our own arithmetic on the list prices, not a measured result.

We assumed a 100,000-token prompt and a 2,000-token reply. With no caching, that costs about $0.0742. If 90,000 of the prompt tokens are cache hits, it costs about $0.0175, roughly 76% less. The sum assumes the list prices apply and that cache hits are billed as listed.

The licence has one catch. Mistral calls it the Mistral Ecosystem Open-Weight License, a modified MIT licence. Commercial use, changes and fine-tuning are allowed, but a business that operates a Model-as-a-Service offering needs a separate commercial licence from Mistral. Because that condition limits some uses, this is open-weight, not open source.

What we do not know yet

Mistral also says Large 4 scores about the same as GLM-5.3, a rival open-weight model, on the Artificial Analysis Coding Agent Index, and uses 47% fewer tokens per task on average. Its page does not say what the 47% is measured against. Both figures are Mistral’s claim alone. We could not find Large 4’s score or tokens-per-task figure for that index on Artificial Analysis’s pages, and the Hugging Face launch post quotes a different score for the same index. Treat it as unchecked.

Other gaps:

• What the 47% is measured against. Mistral names no baseline for “fewer”, so we cannot say whether it means fewer than GLM-5.3, an earlier version of Large 4 or something else.
• Which prices apply where. Mistral does not say if the three prices cover its own API, OpenRouter, Microsoft Foundry or NVIDIA’s endpoints, or whether other hosts charge the same.
• How caching works. We do not know how long a prompt stays cached or what counts as a cache hit. The 76% saving only appears if your prompts qualify.
• Cost of running it yourself. We found no figure for self-hosting, which depends on hardware.
• Benchmark scores. We left them out. Mistral’s page, the Hugging Face launch post and Artificial Analysis give different numbers, and we could not check the outside tests behind some of them.