Mistral Large for an EU Stack: $0.50/$1.50, and the First Model to Fail Our Battery Twice

If your data can't leave the EU, the shortlist is short. Mistral is French, offers EU-hosted inference, and after a large price cut sits at $0.50 input / $1.50 output per million tokens for Mistral Large 3 — down from $2/$6 for Large 2.

That's cheaper on output than every model we've tested this month. The question is what you give up.

We ran Mistral Large through the same four-task battery as everything else, against two non-EU budget models.

It failed two of four

Across a month and roughly twenty configurations, this is the first model to fail more than one task in this battery. Both failures are genuine — we checked the raw output, having been caught by our own graders more than once in this series.

On letter counting it answered 8; the answer is 10. A bare wrong number, no working shown.

On the constraint task it produced four sentences, avoided the forbidden word correctly, and got three of them to exactly five words:

Drops fell from heavy clouds. (5) Puddles formed on wet pavement. (5) Umbrellas opened against sudden downpour. (5) Children splashed through fresh muddy streams. (6)

Close, and wrong. That's the failure mode that matters if you're generating anything with a hard format rule.

It passed the two tasks most representative of ordinary production work — finding a subtle interval-merge bug, and converting messy text into a strict JSON schema. So this isn't "Mistral Large can't code."

Cost

Mistral Large was 2.4–2.8× cheaper than every other model here, despite being by far the most verbose — 900 output tokens against Gemini's 414 and Medium's 407. At $1.50 per million output tokens, verbosity is nearly free.

That is a genuinely aggressive price. If the two tasks it failed aren't in your workload, it's the cheapest capable model we've measured.

How to think about the trade

EU residency is a constraint, not a preference. If GDPR, sector rules or a customer contract require it, the comparison isn't Mistral against GPT-5.6 — it's Mistral against not shipping. At that point a model that passes the bug-finding and JSON tasks at $0.50/$1.50 is a strong answer.

Match it to the failures. It passed structured extraction and subtle-bug detection. It failed exact counting and strict output formatting. If your pipeline needs hard format guarantees, either validate aggressively or pick differently — and note that validation costs nothing and works at any tier.

Mistral Medium passed all four in the same run, at $0.003517 for the battery — more than double Large's cost, because mistral-medium-latest resolves to Medium 3.5 at $1.50/$7.50, three times Large 3's input rate and five times its output rate.

That inversion is worth pausing on: at Mistral, Medium is the expensive one. Large 3 was repriced down aggressively (Large 2 was $2/$6), while Medium 3.5 sits well above it — the older Medium 3 was $0.40/$2.00, so the assumption that "medium is the budget tier" is wrong here in both directions. We verified the alias against Mistral's own /v1/models endpoint rather than trusting the naming.

If you need EU hosting and the format-sensitive tasks matter, Medium is the variant to evaluate — just don't assume it's the cheap one.

EU hosting is not the same as EU headquarters. Several US providers offer EU regional endpoints, and depending on your legal position that may satisfy the requirement at a different price and capability point. Worth confirming what your obligation actually says before narrowing the field.

The broader point

Twenty-odd configurations into this series, almost everything passes almost everything. Mistral Large is the first clear exception, and it's also the cheapest option on the board. That's not a coincidence to hand-wave — it's the first evidence we've collected that the price floor and the capability floor might finally be the same floor.

One model, two tasks, single trials. We'd want considerably more data before calling that a trend. But it's the first crack in a month of uniform results, and worth watching.

Try it yourself

• The counting task — the answer is 10
• The constraint task — four sentences, five words each

To compare EU and non-EU models on your own prompts, that's Promptster.

Run 2026-08-11 via the Promptster API, max_tokens: 4000, one trial per model per task, all in the same session. Mistral Large tested as mistral-large-latest ($0.50/$1.50) and Medium as mistral-medium-latest ($1.50/$7.50); both aliases and prices were verified against Mistral's own /v1/models endpoint on 2026-08-12. Costs computed from list pricing against actual token usage. Single trials are enough for pass/fail against deterministic answers but not to characterise variance — we've measured 26% run-to-run cost variance elsewhere in this series. Both Mistral Large failures were confirmed by reading the raw output rather than trusting the grader.