Mistral API Free Tier Limits: What the Experiment Tier Covers (2026)

Mistral occupies an unusual position among AI providers: it ships open-weight models and a hosted API, which means the question 'is it free?' has two different answers depending on which one you mean. The weights you can run yourself. The hosted endpoints on La Plateforme have a free entry tier that is designed for evaluation, gated by a verification step, and governed by terms that are not identical to the paid tier's.

That split is the source of most of the confusion. Teams read that Mistral is open and conclude the API must be unmetered; teams that start on the free endpoints assume the limits scale linearly into paid and size their architecture accordingly. Neither holds. This guide covers what the free tier actually constrains, how to tell a free-tier rejection from an incident on api.mistral.ai, and the specific signals that mean you have outgrown it.

The practical framing: Mistral's free tier is an evaluation instrument. It is sized to let you decide whether the models fit your workload, not to let you serve customers. Everything below follows from that intent.

The Verification Gate Is Part of the Limit

Unlike providers that hand out a key on signup, access to the free experiment endpoints is gated behind account verification — typically phone verification — before the key does anything useful. This trips up two situations in particular: automated onboarding in CI, and teams in regions where the verification path is awkward. If your first call to api.mistral.ai fails immediately with an authorization-shaped error rather than a quota-shaped one, check the account state before you check your code.

It also has a consequence people miss: because verification is per-account and limits are per-account, the standard workaround of 'just make another key' does not create capacity. It creates an account you now have to verify, on terms that generally prohibit exactly that pattern.

Budget the gate into any evaluation timeline. It is minutes, not days, but it is a human step in a process teams expect to be entirely self-serve, and it reliably surprises the person doing the spike at 6pm.

Free Limits Are Not a Scaled-Down Copy of Paid

The free tier constrains requests per second and per month, and it constrains which models you can reach. The important part is that the ratios between those constraints are different from the paid tier's, so a workload that fits comfortably inside the free allowance can hit a completely different binding constraint the moment you start paying — and vice versa. Sizing your architecture from free-tier behaviour is therefore a mistake with a delayed cost.

The second difference is the model catalog. Mistral's line spans small, fast models through to the large flagship ones, and free access is not uniform across it. An evaluation that only ever ran against a small model has told you very little about the latency, cost or limit behaviour of the model you actually intend to ship, because the constraint profile of the two is not the same.

The third is context. Mistral's models support long contexts, and long contexts are how you empty a token allowance without noticing: a per-request budget that looks generous in isolation becomes a per-minute problem the moment you have a few concurrent long-context calls in flight. Measure tokens per window, not per request.

Free-Tier Rejection or Actual Outage?

From inside your application the two are indistinguishable — a spike in failures, latency climbing, an error rate that looks like something broke — and they demand opposite responses. Back off if it is your allowance; route away if it is theirs.

Read the error body before anything else. A quota rejection names the limit it enforced and usually carries a reset hint; an incident produces generic upstream failures, timeouts and 5xx responses with nothing specific about your account. Then correlate with your own traffic: your limit tracks your peaks precisely, an outage is indifferent to what you are sending. And check a second model on the same key, because a small model still answering while the large one fails points at capacity allocation rather than an outage.

The failure case here is worse on a free tier than a paid one, because there is no support path to escalate down. You cannot open a ticket and get an answer inside your incident window, so the only source of truth about whether api.mistral.ai is healthy is a probe that runs independently of your traffic.

The Open-Weight Escape Hatch

Mistral is the one provider in this comparison where hitting the free-tier ceiling has a third answer beyond 'pay' or 'leave'. Several models are available as open weights, which means the workload that is exhausting your hosted allowance can sometimes move to hardware you control — your own GPU, a rented one, or a third-party host serving the same weights.

Be honest about what that trade actually is. You are exchanging a metered API bill for infrastructure you now operate: capacity planning, model updates, availability, and the on-call rotation that comes with all three. For batch and background workloads with predictable volume, that maths often works. For latency-sensitive user-facing traffic with spiky demand, it usually does not, and teams that make the switch for cost reasons frequently discover the hosted endpoint was the cheaper option once engineering time is counted.

The version of this that most often pays off is hybrid: keep the interactive path on the hosted API where availability is someone else's problem, and move the bulk background work — classification, extraction, embeddings over a corpus — to weights you run. That frees the hosted allowance for the traffic that genuinely needs it.

When to Stop Being Free

The free tier has done its job when you can answer three questions with numbers: does the model quality fit, what does the workload cost per unit at your real volume, and which limit binds first at that volume. If you have those answers, staying free is now costing you rather than saving you.

The hard triggers are the same everywhere and worth stating plainly. Someone outside your team depends on the response. Rejections appear in normal traffic rather than only at peaks. You are engineering around the ceiling at a developer-hours cost that exceeds the invoice. Or you are about to send data whose handling requirements do not match the terms attached to free usage — which on any provider is the one that should end the conversation immediately.

Move before the launch, not during it. And keep independent monitoring on api.mistral.ai either way: paying removes the ceiling, not the incidents, and after the upgrade an outage is rarer and therefore catches you less prepared.

Frequently Asked Questions

Is the Mistral API free?

There is a free entry tier on La Plateforme intended for evaluation rather than production, and it is gated behind account verification before the key will do anything useful. Separately — and this is where most of the confusion comes from — Mistral publishes open weights for several models, so 'free' can also mean running them yourself on your own hardware, which is a completely different cost structure with completely different failure modes. The hosted free tier has an allowance and no meaningful service commitment; the self-hosted route has no allowance and no service commitment at all, because the service is now you.

Why does my Mistral free API key not work at all?

Most often because the account has not completed verification. Free-tier access on La Plateforme is gated behind a verification step, typically phone verification, and until it is done the key exists but calls fail in an authorization-shaped way rather than a quota-shaped one. Check the account state in the console before you debug your client. This catches teams doing automated onboarding in CI especially hard, because it is a human step in a flow they reasonably expected to be entirely self-serve.

What are the Mistral free tier rate limits?

They are enforced per second and per month, published per model, and they change as capacity and the catalog change — so read them from the console for the exact model string you are calling rather than trusting any article, including this one. The more useful point is that the ratios between the free limits are not the same as the ratios between the paid ones, so a workload sized against free-tier behaviour can hit a completely different binding constraint after you upgrade. Do not use free-tier numbers to size a production architecture.

Is it cheaper to self-host Mistral than to use the API?

Sometimes, and the answer depends almost entirely on utilisation rather than on token price. Self-hosting trades a metered bill for hardware you keep running whether or not requests arrive, plus the operational load of capacity planning, model updates and availability. For predictable high-volume batch work, that trade frequently wins. For spiky user-facing traffic, the hosted endpoint is usually cheaper once engineering time is honestly counted. The hybrid version wins most often: interactive traffic hosted, bulk background work on weights you run.

How do I tell a Mistral rate limit from an outage?

Read the error body: a quota rejection names the limit and usually carries a reset hint, whereas an incident produces generic upstream errors, timeouts and 5xx responses with nothing account-specific in them. Then correlate the failures with your own request rate, because your limit tracks your peaks and an outage does not. Try a second model on the same key as a tiebreaker. On a free tier this matters more than on paid, because there is no support path you can escalate down inside an incident window — independent monitoring of api.mistral.ai is the only source of truth you control.

Are free-tier prompts used to improve Mistral's models?

Free and paid usage are commonly governed by different terms on retention, review and service improvement, and the free-tier terms are the ones people skip. Read the current agreement for the tier you are actually on before sending anything regulated or customer-identifying. The durable rule is to decide what class of data your evaluation may touch before you start, because free-tier prototypes have a way of accumulating real customer data long before anyone revisits the terms, and migrating out of that position later is far more expensive than starting on paid.

When should I move off the Mistral free tier?

Once the free tier has answered the three questions it exists to answer — does the quality fit, what does it cost per unit at real volume, and which limit binds first — it has stopped earning its place. Move immediately if someone outside your team is waiting on the response, if rejections show up in ordinary traffic rather than only at peaks, if you are spending more on engineering around the ceiling than the invoice would be, or if the data you are about to send does not match the handling terms attached to free usage.

Related Guides

Free Tier Limit, or Is Mistral Actually Down?

On a free tier there is no support path to escalate down inside an incident window. API Status Check probes api.mistral.ai independently of your traffic and alerts on errors and latency, so you can tell your own ceiling from their outage in seconds.

Start Your Free Trial →