Mistral Large 4 gets updated weights and lower token pricing on Vercel AI Gateway

Mistral Large 4 now runs on updated weights across Vercel's AI Gateway, and the change arrives alongside reduced pricing for input, cached input and output tokens. According to the Vercel blog, developers already calling mistral/mistral-large-4 through the gateway need do nothing: the existing model ID resolves to the new weights, and the lower prices apply automatically.

What the updated weights change

Vercel attributes the substance of the update to Mistral's release announcement. The refreshed weights reportedly improve agentic coding and defensive cybersecurity performance, and deliver greater token efficiency on coding tasks, meaning the model should consume fewer tokens to complete comparable coding work, which compounds with the per-token price cut.

The announcement also lists improvements to spreadsheet creation and editing, web research, and resistance to prompt injection. The injection-resistance claim stands out for agent use cases, since models that browse the web or ingest tool output are routinely exposed to untrusted text.

Pricing on the gateway

The Vercel blog confirms lower prices for input, cached input and output tokens but does not include specific per-token figures in the changelog entry. Vercel reiterates that AI Gateway mirrors provider pricing with no markup and charges no platform fee on inference, including on Bring Your Own Key (BYOK) requests, so reductions flow through to gateway users rather than functioning as a platform-side discount.

No migration required

Because the identifier is unchanged, the rollout is effectively silent for existing users. The changelog entry does not mention a separate alias or snapshot for the previous weights, so teams that rely on pinned model behaviour for evaluations may want to re-baseline before assuming nothing else has shifted. New capabilities and lower cost arrive through the same string.

Running it inside coding agents

For use in fx, Pi, Claude Code, Codex, Cursor or other harnesses, Vercel says to install the latest Vercel CLI and run npx vercel ai-gateway setup, then select mistral/mistral-large-4 inside the agent. The model is also available in Vercel's model playground for direct testing without writing any code.

What AI Gateway handles

Vercel positions AI Gateway as a unified API for calling models, tracking usage and cost, and configuring retries, failover and performance optimizations that the company says deliver higher-than-provider uptime. The service also offers built-in custom reporting, Zero Data Retention support, budgets tied to API keys, and routing rules across its model catalogue.

Why it matters

Two things make this more than a routine version bump. First, the economics: better token efficiency on coding tasks stacked on top of lower per-token pricing is a double reduction for exactly the workloads where token volume drives most of the bill, namely coding agents running long sessions. Second, the delivery mechanism: shipping new weights under an unchanged model ID means teams pick up the improvements and the savings by default, with no code change, though it also removes an easy way to stay on the old behaviour. Improved prompt-injection resistance is a timely addition as more of these models get wired into agents that read the open web, where untrusted content is the norm rather than the exception.