Mistral API Free Tier: Limits and a Prepaid Alternative

What to expect from the Mistral API free tier

Most people who search for the mistral api free tier want to know three things: how many requests they get, which models are included, and what happens when they outgrow it. Mistral decides all three, and the numbers are the kind that get adjusted, so any figure copied into an article like this one would be stale soon. Open their documentation and billing pages and read the current limits directly.

What you can predict is the shape. Free allowances across the industry tend to share a few traits: lower rate limits than paid plans, a cap on tokens or requests per period, terms about how prompts may be used, and no promise of capacity during busy hours. None of that is a criticism. A free tier is a marketing and testing tool, and it works fine for that job.

It becomes a problem when a prototype quietly turns into a product. The day your app has real users, you want predictable limits, a balance you can see, and an answer to the question "what happens if the model refuses a perfectly lawful request?" The rest of this page covers how to plan for that moment.

A checklist for reading any free tier's terms

Before you build anything on a free allowance, answer these questions from the provider's own pages. Write the answers down with the date, because they change.

1. What is the request limit? Look for requests per minute, tokens per minute and any monthly cap, as separate numbers.
2. Which models are covered? Free access sometimes applies to only some models.
3. Can the data be used for training? Free usage on some services comes with different data terms than paid usage. Read the privacy section, not the pricing section.
4. Is commercial use allowed? Some free plans are for evaluation only.
5. What is the upgrade path? Find out what a paid account costs and whether limits rise immediately.
6. What are the content rules? If your use case is fiction, roleplay or other adult-oriented writing that is lawful, check how the model handles it before you invest time in prompts.

The last item is the reason many people end up on a different service. A model with built-in refusals can pass your first ten test prompts and then decline the eleventh, and a free tier gives you little room to experiment around that.

The trial here, in numbers

This API is independent of Mistral and offers a small trial instead of a permanent free tier. New users sign up with Google or with an email and password, then click once on the API key page to claim $0.50 of credit. There is no card step. It is valid for 7 days and limited to one per person.

The trial key runs under lower limits than a funded one: 2 concurrent requests and 60 requests per minute. That is plenty for testing a script, a SillyTavern-style chat client or a small batch job, and not enough for a public app.

What $0.50 buys

Token prices are $0.15 per 1M input and $0.50 per 1M output. Some concrete budgets:

• All input, no output: $0.50 / $0.15 per 1M is about 3.33M tokens of prompt.
• All output: $0.50 / $0.50 per 1M is exactly 1M tokens of generated text.
• A typical small request of 1,000 tokens in and 300 out costs 0.00015 + 0.00015 = $0.0003. So $0.50 covers about 1,666 of them.
• A long-document request of 20,000 tokens in and 1,000 out costs 0.003 + 0.0005 = $0.0035, so about 142 of them.

In practice that is enough to run a few hundred real prompts from your own use case, which is exactly what an evaluation needs. If the answers are not what you want, you have lost 50 cents and no card was ever on file.

Moving from trial to production

When the trial works, the step up is a prepaid top-up of any whole amount from $20 to $500. There is no subscription and the credit never expires, so topping up the minimum and leaving it alone for a month costs nothing extra. Larger amounts add a bonus: +5% from $50, +10% from $100, +15% from $200. A $200 top-up therefore gives you $230 of credit.

A funded key raises the limits to 8 concurrent requests and 300 requests per minute. At the small-request size from above ($0.0003 each), a $20 top-up covers about 66,000 requests: 20 / 0.0003 is 66,666. For the full rate card, see the pricing page.

Paying is by crypto (USDT on TRC20, USDC on Base).

The difference you will feel most is predictability. Instead of wondering whether a shared free allowance is throttled today, you have a balance in dollars, a header on every response that shows it, and two documented limits. A failed or refused request is not billed, so a bad day on the server side does not eat your credit.

Code: run the trial and watch the balance

The API follows the OpenAI chat format, so your existing SDK works with a new base URL and key. This script sends one request, prints the answer and reads the remaining balance from the X-Balance-USD header.

Copyfrom openai import OpenAI

client = OpenAI(
    base_url="https://api.mistralapi.com/v1",
    api_key="YOUR_API_KEY",
)

raw = client.chat.completions.with_raw_response.create(
    model="uncensored",
    messages=[{"role": "user", "content": "Give me three names for a fantasy tavern."}],
    max_tokens=200,
)
resp = raw.parse()

print(resp.choices[0].message.content)
print("tokens in/out:", resp.usage.prompt_tokens, resp.usage.completion_tokens)
print("balance left: $", raw.headers.get("X-Balance-USD"))

Run it a few times and you will see the balance move by fractions of a cent. That is the best way to build a feel for what your prompts really cost before you commit to a top-up. If you want the same thing from the shell:

Copycurl -i https://api.mistralapi.com/v1/chat/completions   -H "Authorization: Bearer YOUR_API_KEY"   -H "Content-Type: application/json"   -d '{"model":"uncensored","messages":[{"role":"user","content":"Say hello."}],"max_tokens":20}'

The -i flag prints the response headers, including X-Request-Id, X-Balance-USD and the two X-RateLimit-Limit headers. Keep the request ID if you ever contact support.

Handling rate limits without losing requests

Limits exist on every plan, free or paid, so write the retry logic once. This API answers with HTTP 429 in three situations. The error code rate_limited means you exceeded requests per minute. The code concurrency means too many requests were in flight at once. A 429 with a Retry-After: 5 header means temporary overload. In all three cases the right response is to wait and send the request again.

Copyimport time
from openai import OpenAI, RateLimitError

client = OpenAI(base_url="https://api.mistralapi.com/v1", api_key="YOUR_API_KEY")

def ask(prompt, tries=5):
    for attempt in range(tries):
        try:
            r = client.chat.completions.create(
                model="uncensored",
                messages=[{"role": "user", "content": prompt}],
            )
            return r.choices[0].message.content
        except RateLimitError:
            time.sleep(5 * (attempt + 1))
    raise RuntimeError("still rate limited after retries")

On the trial key, keep your worker pool at two or fewer. A streamed request holds its concurrency slot until the stream finishes, so two long streams can block a third request even if you are far below 60 per minute.

A realistic evaluation plan for the trial

Half a dollar goes a long way if you spend it on the right tests. A plan that works well:

1. Collect 100 to 200 real prompts. Use actual user messages, support tickets or story starters from your project, not invented examples. Include the awkward ones that another model might decline.
2. Run them once with fixed settings. Use a low temperature so you are judging the model rather than the dice.
3. Read a sample by hand. Twenty replies is enough to spot a pattern. Note any that you would not ship.
4. Measure tokens. Average the prompt and completion counts from usage.
5. Project the monthly bill. Multiply the per-request cost by your expected volume.

For scale: 200 prompts at 800 tokens in and 400 out use 160,000 input tokens (0.024 dollars) and 80,000 output tokens (0.04 dollars), so one full pass costs $0.064. You could run that test set about seven times inside a $0.50 trial, and the 7-day validity is rarely the limit. Changing the system prompt between passes is the cheapest tuning you will ever do.

Stop the trial early if the replies are not good enough for your task. Nothing renews, and no card was entered, so there is nothing to cancel.

Which route fits you

Use a free tier when you are checking whether a model is capable enough, when you are writing a tutorial, or when the volume is a handful of requests a day. Use a prepaid key when other people depend on the output, when you need to know your limits in advance, or when your use case sits where a stricter model might refuse. The quickstart covers the switch in a few lines if you are coming from an existing Mistral integration.

Keep one rule in mind whichever you choose: the service does not generate sexual content involving minors, and no tier or setting changes that.

More guides

• Mistral AI API Pricing vs an Uncensored Pay-Per-Token API
• Uncensored Mistral API: Pay Per Token