Mistral's documentation shows Mistral Large 4 at $0.68 input, $0.07 cached input and $2.09 output per million tokens, with $1.36, $0.14 and $4.18 struck through. Run your evaluations now, but build the business case on the list price.
Mistral opened a public preview of Mistral Large 4 on October 6, and the first thing worth noticing is on its documentation page rather than in the launch post. The model page shows a price of $0.68 per million input tokens, $0.07 per million cached input tokens and $2.09 per million output tokens, sitting next to struck through prices of $1.36, $0.14 and $4.18. The launch announcement quotes only the higher pair, $1.36 in and $4.18 out. In practice, anyone testing the model today is paying half of what the list says.
That is a gift for evaluation and a trap for planning. A pilot that looks cheap at the current rate can double in cost the day the discount ends, and nothing on the page says how long it lasts. The fix is simple: run as many evaluations as you can while the rate is low, and model every production cost estimate at $1.36 and $4.18. If the cheaper rate turns out to be permanent, you will have built in headroom rather than a budget overrun.
The documentation also lists a 1 million token context window, function calling and structured outputs on both the chat completions and conversations endpoints, and batching on the batch endpoint. Batching is the natural home for a large evaluation run, because you can queue a full test set without tying up interactive capacity, and structured outputs let you score results automatically instead of reading them by hand.
One code level detail can break a pipeline quietly. Mistral's reasoning guide describes two reasoning_effort settings for its models: high, which returns a full thinking chunk before the final answer, and none, which omits it. With high, message.content comes back as a list of chunks, a thinking chunk and a text chunk, rather than a plain string; with none it is a plain string. Any code that assumes a string, such as calling strip on the content or writing it straight to a database, needs to check the type first.
Finally, remember the bigger option on the horizon. Mistral says it will release the weights by the end of October. If your evaluations go well, the decision you are really making is not only which API to call but whether to self host, so use the preview window to collect the quality and cost data that decision will need.
Try it today
- Run your evaluation set now through the batch endpoint while the documentation shows the discounted rate.
- Model production costs at the struck through list prices of $1.36 input and $4.18 output per million tokens, not the current rate.
- Use structured outputs so evaluation results can be scored automatically.
- If you set reasoning_effort to high, handle message.content as a list of thinking and text chunks; with none it is a plain string.
- Revisit the self hosting question when Mistral releases the weights, which it says will happen by the end of October.