Mistral Large 4 gets new weights on AI Gateway: same model ID, lower input, cached-input and output prices

Mistral Large 4 on AI Gateway now runs updated weights, and input, cached-input and output token prices are lower at the same time, so callers using the same model ID get the new version directly. Previously that model ID pointed to the older weights and older prices; now the same ID points to new weights and new prices, and callers do not need to change the model string. The change is described in Mistral's release announcement, which lists both the capability changes from the new weights and the scope of the price cut.

The new weights target agentic coding and defensive cybersecurity, use fewer tokens on coding tasks, and improve spreadsheet creation and editing, web research, and resistance to prompt injection. These are capability changes relative to the prior version, described by the announcement as directions of improvement rather than specific scores. The basis is the capability description in Mistral's release announcement; the announcement gives no per-item benchmark numbers.

The report is limited to what Mistral's release announcement states: the weight update and price cut for mistral/mistral-large-4 on AI Gateway, plus the directions of capability improvement the announcement lists; the announcement gives no per-item benchmark scores.

The announcement describes capability changes as directions of improvement and gives no specific evaluation scores; the size of the price cut is also not given in the material.