Cost control: the cheapest way to get the result from Mistral

Mistral is already one of the cheaper providers, which makes it easy to get lazy. The waste does not come from the list price; it comes from using Large where Small would do, re-sending context, paying real-time rates for overnight jobs, and keeping a subscription you stopped needing. This lesson is the checklist.

Rule 1: route, do not default

Build a small test set of twenty real examples for each task. Run it on Ministral 3 8B, then Mistral Small 4, then Mistral Large 3. Use the cheapest model that passes. Most extraction, classification and formatting tasks pass at Ministral or Small. Only reasoning-heavy or nuance-heavy tasks need Large, and Medium 3.5 needs to prove it is worth five times the output price of Large 3.

Rule 2: batch anything that can wait

If a job runs overnight, on a schedule, or over a queue, use the batch API. It is priced below real-time; the current discount is on the official pricing page. Nightly summaries, weekly reports and back-fills all qualify. Interactive chat does not.

Rule 3: trim input, cap output

• Send the section, not the document. Retrieval plus a 4K-token chunk beats a 200K-token dump on both cost and accuracy.
• Reset long conversations. Each turn re-sends history; start a new thread with a two-line summary once the chat passes twenty turns.
• Set max_tokens and say "no preamble". Output is the expensive side on most Mistral models.
• Cache your own results. If the same document gets the same question twice, you should be reading a database, not paying twice.

Rule 4: consider self-hosting at steady volume

Because Small, Ministral and Nemo are open-weight, you can run them on your own GPUs. Do the arithmetic honestly: monthly API spend on that model versus GPU rental plus an engineer's time. Self-hosting usually wins only above steady, predictable volume or when a data-residency rule forces it. Below that, the API is cheaper. Mistral's EU hosting often satisfies the residency requirement without self-hosting at all.

When to downgrade or cancel

• You have a paid Le Chat tier and cannot name the feature or limit that made you buy it. Downgrade to Free for a month and see if anything breaks.
• You pay for Le Chat and for the API. One of them is almost certainly idle. Keep the API if you build, keep Le Chat if you chat.
• Your API bill is dominated by Large or Medium 3.5 on tasks you never tested on Small. Run the test set.
• Your bill is flat while your usage fell. Subscriptions do not notice; you have to.