As of October 2026, the default state for a new Mistral account is Free mode: API access is enabled by default and no credit card is required, but the official documentation also states that "usage and rate limits apply." The implication is direct: the free tier is good for validation, but it should not be assumed sufficient for production.

This article breaks the free tier into four questions: where the quota comes from, where to find the rate limits, what the data terms say, and when an upgrade becomes necessary. Specific usage allowances, model coverage, and data rules may change; the console and the contract take precedence.

Note: Mistral, Le Chat, Codestral, and other names are trademarks of their respective owners. This article is for technical explanation only and does not imply affiliation, agency, or endorsement. Plans, quotas, and data terms may change; refer to the Mistral official console, documentation, and contract.

1. What the free tier is: enabled by default, but not unlimited trial

Mistral currently organizes plans into Free mode, Pro, Education, Team, and Enterprise, and states that plans are global, applying simultaneously to Vibe, Studio, and the API. New accounts start in Free mode, API keys are created in the Studio console, and the official quickstart explicitly says that "API access is enabled by default with no credit card required."

What is free is the subscription fee, not the resources. The public documentation only says that usage and rate limits apply; it does not list fixed values. The actual limits must be checked in the console under Admin Panel › API › Limits and on the workspace usage page. They may differ by model, time, and workspace, so hard-coded "free quota numbers" in tutorials are only a reference.

2. How Free mode, paid plans, and Enterprise differ

Dimension Free mode Paid plans (Pro / Team, etc.) Enterprise
How to start Enabled by default for new accounts, no card required Subscription plan, with optional prepaid pay-as-you-go Contact the Mistral account team to enable
Usage and rate limits Usage and rate limits apply; values in the console Monthly usage included; overage needs prepaid balance Defined by contract; dedicated quota negotiable
Behavior after overage May stop until the next billing period Returns 402 when balance is low; restored after top-up Governed by contract and service level
Data used for training May be used for training; API toggle can opt out Can opt out of training, subject to terms Official materials state data stays in the customer environment; confirm in contract
Zero data retention (ZDR) Not available on the free tier Available on paid plans by application and review Can be written into an enterprise contract
Support and service level Official support channels, no enterprise SLA Pro includes chat and email support Custom SLA, dedicated support, private deployment

This table describes where the rules live, not the specific allowances. When selecting, collect console screenshots, contract terms, and ticket replies, not numbers from someone else's article.

3. Six limits of the free tier

  • Rate and concurrency: requests per second, tokens per minute, and tokens per month may all have ceilings, and limits are calculated per workspace and shared by every key in that workspace. Exceeding them commonly results in 429 responses or queuing.
  • Shared monthly usage: the monthly allowance within the plan also supplies Studio, API, and Vibe Code. A stress test in one scenario can consume the quota needed by the others.
  • Model coverage: whether a model is callable on the free tier depends on the console model list. Long context, tool calling, JSON output, vision, and OCR capabilities also need item-by-item verification; "it can answer" is not enough.
  • Data terms: API inputs and outputs may be included in the training program, and the default state and toggle location are described in the next section. The free tier is also outside the scope of zero data retention.
  • Service level: enterprise SLA, dedicated support, and private deployment are Enterprise contract items. Do not design a business around an availability commitment on the free tier.
  • Quota and credit expiry: monthly usage resets by billing period. If the account holds granted or purchased credit, the Credit record in the console shows the source, balance, and expiry date. Promotional rules may change, so leave time for migration before expiry.

4. Data terms: the most overlooked part of the free tier

The Mistral help center states that API inputs and outputs may be included in the model training program, and that users may opt out at any time. The Studio/API toggle is under Admin Panel › Privacy, where "Anonymous improvement data" can be turned off. Vibe has a separate privacy toggle, and neither replaces the other.

Two details are often missed. First, the DPA states that Mistral may use customer data to train models as an independent controller unless the customer has opted out; feedback such as thumbs up or thumbs down, together with the associated inputs and outputs, is still used for training even after opting out. Second, ZDR and training opt-out are two independent controls: ZDR is available only on paid plans, requires an application and review each time, covers only supported stateless endpoints, and excludes Labs models.

For scenarios involving contracts, code, customer personal information, or unpublished business data, do not use the free tier to "get it running first." Either complete data classification and redaction first, or go directly to a paid plan with ZDR, or choose private deployment and write the data-processing boundary into the contract.

5. Four kinds of work the free tier suits

First, SDK and network connectivity validation: confirm that the key, base URL, timeouts, proxy, and error-code handling all work. Second, model selection: compare output quality, latency, and stability on the same batch of tasks. Third, internal demonstrations: show a workflow to colleagues without touching real customer data. Fourth, usage sampling: accumulate token-consumption samples for a future paid budget instead of guessing.

The unsuitable scenarios are equally clear: customer-facing production endpoints, peak concurrency, long-running agents, internal systems that need a stable SLA, and any business that cannot accept data entering a training pipeline. The free tier is a validation tool, not a production solution.

6. One-week validation checklist

  1. Create a separate key with an expiry date, record its creation date and purpose, and do not mix it with long-lived keys.
  2. Sample 20 to 30 tasks from real business, covering simple, medium, and complex levels, with fixed inputs and scoring criteria.
  3. Run one round during the morning peak and one at night, recording 429s, 5xxs, timeouts, time to first token, and p95 latency.
  4. Separately validate critical capabilities such as long context, JSON output, tool calling, vision, or OCR; do not test only ordinary question answering.
  5. Confirm that the target model is available in the current workspace, and record the model name and version identifier.
  6. Turn off "Anonymous improvement data" for the API and confirm the state has been saved; if Vibe is also used, check the Vibe toggle separately.
  7. Simulate quota exhaustion: observe whether requests fail, queue, or stop until the next billing period, and design degradation and alerts accordingly.
  8. Calculate average daily tokens and peak usage over a week, compare them with the free tier's actual available capacity, and estimate when an upgrade is needed.

The script below writes the status, tokens, and latency of each call to JSONL, so the results can be summarized directly after a week:

import json
import time
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_MISTRAL_API_KEY",
    base_url="https://api.mistral.ai/v1",
)

def probe(prompt: str) -> dict:
    start = time.time()
    try:
        resp = client.chat.completions.create(
            model="mistral-small-latest",  # example model name; check the console list
            messages=[{"role": "user", "content": prompt}],
            max_tokens=256,
            timeout=60,
        )
        row = {
            "status": 200,
            "input": resp.usage.prompt_tokens,
            "output": resp.usage.completion_tokens,
            "latency_ms": int((time.time() - start) * 1000),
        }
    except Exception as exc:
        row = {"status": "error", "error": type(exc).__name__}
    with open("mistral-probe.jsonl", "a", encoding="utf-8") as f:
        f.write(json.dumps(row, ensure_ascii=False) + "\n")
    return row

probe("In three sentences, what should I watch for in a free API trial?")

7. Upgrading from the free tier to a paid plan

Mistral's pay-as-you-go billing is prepaid, and an organization must have credit before it can be enabled. Under the billing rules, the monthly allowance within the plan is consumed first; after that, both pay-as-you-go being enabled and a positive balance are required. When the balance is zero, requests return HTTP 402 and access resumes after a top-up. When paid-plan credit reaches 80 percent, email reminders go to administrators and billing contacts.

Before upgrading, do three things: split production, staging, and development into separate workspaces, each with its own monthly spending cap; set an organization-wide cap plus a smaller cap for each workspace, so that one workspace hitting its limit only stops itself; and write handling for 402, 429, and 5xx into the code, with backup models or channels ready.

If the team also needs RMB settlement, consolidated multi-model billing, or optimized access from mainland China, the ZSQH AI Gateway can serve as a parallel channel. It is an OpenAI-compatible interface, so migration mainly involves changing the base URL, key, and model name. Specific models and prices are listed on the model pricing page, and an account can be created directly via sign-up. An extra channel is not an extra layer of insurance; degradation drills are still required.

Related reading: DeepSeek API enterprise procurement checklist, how to choose a multi-model API platform, API relay usage guide.

8. Frequently asked questions

Does the free tier require a credit card?

The official quickstart states that API access is enabled by default and no credit card is required. But free access does not mean there are no limits; usage and rate limits still apply, and the specific values are shown in the console.

Does the free quota expire?

The monthly allowance within the plan resets by billing period. Granted or purchased credit may have its own validity period, and the Credit record in the console shows the expiry date. Promotional credit and trial policies may change; do not treat them as a long-term budget source.

Are free-tier inputs and outputs used for training?

They may be included in the training program. "Anonymous improvement data" can be turned off under Admin Panel › Privacy, and the Vibe toggle must be handled separately. Feedback such as thumbs up and thumbs down, and the associated content, are still used for training even after opting out, so sensitive data should not go through the free tier.

Can the free tier apply for zero data retention?

ZDR is available only on paid plans, requires an application and review, covers only supported stateless endpoints, and excludes Labs models. ZDR and training opt-out are two independent controls and cannot substitute for each other.

Can the free tier go straight into production?

Not recommended. The free tier has no enterprise SLA, and limits and model coverage may change. A production system needs at least a paid plan, spending caps, error handling, monitoring, and a rehearsed backup channel.

Related reading: