Mistral Large 3 Deprecation: Migrate to Large 4 by April 30, 2027 and What It Costs
Mistral deprecated Mistral Large 3 (mistral-large-2512) on October 31, 2026 and will retire it from its API on April 30, 2027, with Mistral Large 4 named as the replacement. If you call Large 3 today, your requests keep working until the retirement date, but the Large 4 list price of $0.70 per million input tokens and $2.10 per million output tokens is 40% above the $0.50 and $1.50 that Mistral's pricing page shows for its Large model.
Key facts
- October 31, 2026: Mistral Large 3 (25.12) was deprecated. Retirement is April 30, 2027. Replacement: Mistral Large 4.
- The same table lists Mistral Medium 3.5 (26.04) with the same two dates and the same replacement, Mistral Large 4.
- Mistral Large 4 (v26.10) is a general-availability, open-weight multimodal model with a 1M-token context window, 1.05T total and 52B active parameters.
- Large 4 API price: $0.70 input, $0.07 cached input, $2.10 output per million tokens. A third-party tracker reports these rates applied from October 29, 2026, after a $1.36 and $4.18 list price seen on October 25.
- Mistral's pricing page lists an unversioned "Mistral Large" at $0.5 input and $1.5 output per million tokens, with batch processing 50% cheaper and cached input up to 90% cheaper.
What deprecated means for your integration
Mistral's model overview keeps a deprecated models table with a deprecation date, a retirement date and a replacement for each model. Deprecation is the warning. Retirement is the date the model stops answering API calls. For Large 3 that is about five and a half months from today, which is enough time for a careful migration and not enough for a quarter-by-quarter wait-and-see.
If you pin the model ID mistral-large-2512 in code, config files or a prompt-management tool, that string is what has to change. Search for it across services, notebooks and scheduled jobs, not only your main application.
What Large 4 costs against the old rate
The step up is 40% on both input and output: $0.70 against $0.50, and $2.10 against $1.50. A workload of 100 million input tokens and 20 million output tokens a month comes to $80 on the Large rates ($50 input plus $30 output) and $112 on Large 4 ($70 plus $42). That is our arithmetic from the published list prices, not a Mistral figure.
One caveat on the baseline. Mistral's pricing page labels its $0.5 and $1.5 line "Mistral Large" without a version number, so check your own invoice to confirm that is the rate you pay for Large 3. Cached input on Large 4 is $0.07, so applications that resend a long system prompt can recover part of the increase.
Mistral Medium 3.5 users face the same migration, since the table points Medium 3.5 at Large 4 too. We did not find a Medium 3.5 price in the pages we checked, so we cannot say what that move costs.
What else changed with Large 4
The Large 4 model card shows a 1M-token context window, structured outputs, function calling, document Q&A, batching and built-in tools. The weights are published under the Mistral Ecosystem Open Weight License. A third-party price tracker reports that this license covers self-hosting, fine-tuning and on-premises use, and does not allow reselling the model as a service without a separate commercial license. Read the license text itself before relying on that summary. The same tracker lists the FP8 weights at about 1,061 GB, so self-hosting is a hardware decision, not a weekend project. For more on when that choice pays off, see our piece on token prices and vendor margins.
What to do before April 30, 2027
- Search every repository, config and job for mistral-large-2512 and list where it runs.
- Replay a sample of real production prompts against Large 4 and compare answer quality, latency and token counts, because output length drives most of the bill.
- Recalculate your monthly cost at $0.70 and $2.10, and add prompt caching if you resend the same long context.
- Check whether a cheaper model covers part of your traffic. Price moves like this one are why we track them, as in OpenAI's 80% cut to its cheapest GPT-5.6 tier.
- Put the April 30 date in your calendar with a cutover target at least a month earlier. Browse the rest of our pricing coverage in the archive.
Questions people ask
When does Mistral Large 3 stop working?
Mistral's deprecated models table gives April 30, 2027 as the retirement date for Mistral Large 3 (25.12). It was deprecated on October 31, 2026.
What replaces Mistral Large 3?
Mistral Large 4. The deprecated models table names it as the replacement for Large 3, and for Medium 3.5 as well.
How much does Mistral Large 4 cost?
The model card lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.
Sources