Mistral Large 4's API: Chat, Agents, Batch and the Endpoints Most Guides Skip

Most write-ups of a new model stop at the chat completion. Mistral's own documentation for Mistral Large 4 lists four product surfaces — chat, agents and conversations, batch, and built-in tools — and the pricing page adds a second axis almost nobody uses: standard, batch and priority inference tiers, plus a regional-inference switch. Knowing which of those you are on is worth more than knowing the model's benchmark score.

The surfaces, as Mistral documents them

• Chat Completions — code>/v1/chat/completions/code>. The OpenAI-shaped endpoint, with structured outputs, function calling, prefix completion and document QnA listed as features.

• Agents & Conversations — code>/v1/agents/code> and code>/v1/conversations/code>. Server-side state rather than client-side message assembly: you keep a conversation on Mistral's side and refer to it, which removes the "re-send the whole history" pattern that dominates input-token bills.

• Batch — code>/v1/batch/code>. Asynchronous job submission, priced on the batch tier rather than the standard one.

• Built-in Tools — the model calling Mistral's own tool set rather than yours.

Those four are not the same product with different URLs. The agents surface changes your cost structure — if conversations live server-side, the input side of your bill stops scaling with history length. The batch surface changes your price tier. Both are documented, and both are easy to miss because every tutorial starts and ends with chat completions.

The tier axis nobody prices

Mistral's API pricing page carries a filter row — Standard, Batch, Priority — above the model grid, and a separate toggle for regional inference and one for currency. That means the $1.36/$0.14/$4.18 card is the standard tier. Batch is a different number. Priority is a different number again, and neither is shown in the default view most buyers screenshot.

Before you compare Mistral Large 4's price against anything, confirm which tier the number you are reading belongs to. The practical asymmetry: a batch workload that lands on the standard tier is paying retail for a capability the price list already discounts.

What the model behind the endpoints is

1.05 trillion parameters with 52 billion active by Mistral's card, on a granular sparse mixture-of-experts with 128 routed experts and four active per token; a 1.6-billion-parameter vision encoder; text and image in, text out; reasoning and text as the class. Context is 1M configured, 524,288 measured on the tested configuration. Weights are public under the Mistral Ecosystem Open-Weight License, 1,060.6 GB in bf16 with a 610.1 GB NVFP4 sibling.

One consequence of that licence deserves a line in any integration plan: clause 2 reserves "Model as a Service" — giving third parties access to the model with meaningful control over inputs, parameters or training data — for a separate commercial agreement. An embedded feature, an internal deployment, and mere relaying to somebody else's endpoint are all carved out. Calling the API is not affected; reselling raw access through your own API is.

The integration order that saves money

• Pick the surface first. If your product holds conversations, use the agents/conversations endpoints and stop re-sending history; that is an input-token line, and this model's input price is $1.36 per million.

• Pick the tier second. Anything asynchronous belongs on batch, and the pricing page will show you the delta once you switch the filter.

• Measure the context ceiling third. 524,288 measured, not 1M claimed — send one oversized request to your host and read the error.

• Add caching last. $0.14 cached input against $1.36 standard is the single largest lever on a steady-state bill, provided your prefix is stable. A timestamp at the top of a system prompt turns it into a full-price call.

Mistral Large 4 is not on OrcaRouter's catalogue, so this is not a route we can offer — it is callable through Mistral's own API and third-party hosts. The general shape we do offer is one OpenAI-compatible endpoint over 224 models with provider list prices passed through at 0% markup, where a vendor's tier and price changes land the day they are announced rather than at the next renewal.