Mistral Large 3 costs $0.50 per million input tokens and $1.50 per million output tokens, read from the Mistral documentation on 16 August 2026, with cached input discounted ninety percent across the standard offering.
The striking part is where that lands: Mistral Medium 3.5, the tier below it, costs $1.50 and $7.50, so the flagship is three times cheaper than its own mid-range on input and five times cheaper on output. No other provider we track is priced that way.
The flagship is the cheap one
Read the table in order of tier and the numbers go the wrong way.
| Model | Input | Output |
|---|---|---|
| Mistral Large 3 | $0.50 | $1.50 |
| Mistral Medium 3.5 | $1.50 | $7.50 |
| Mistral Small 4 | $0.15 | $0.60 |
| Ministral 3 14B | $0.20 | $0.20 |
| Ministral 3 8B | $0.15 | $0.15 |
| Ministral 3 3B | $0.10 | $0.10 |
Mistral Large 3 at $0.50 in and $1.50 out per million, read 16 August 2026
Ministral 3 14B is also dearer than Mistral Small 4 on input, at $0.20 against $0.15, despite sitting below it in the range. The only monotonic stretch is inside the Ministral family itself.
We are not going to guess why. What we can say is what it means for you: on this provider, do not use tier names as a proxy for price. Read the grid.
On Mistral, moving up to the flagship can cut your bill. That sentence is false at every other provider on this site.
What output costs, relative to input
The multiplier between output and input is what decides a generative bill, and Mistral spreads it unusually wide across its own catalogue.
Mistral Large 3 at 3x is the tightest output ratio of any flagship we track: Anthropic runs at five times across the Claude line and OpenAI at five on its current models. On work that writes a lot, that ratio moves your bill more than the input price does.
The Ministral models go further and bill output at exactly the input rate. A one-to-one ratio is close to unique, and it makes those models unusually predictable to budget: your bill scales with total tokens and you do not have to forecast the split.
What a month actually costs
An agent workload of twenty million input and four million output tokens, priced with and without caching.
| Monthly workload | Large 3 / Medium 3.5 / Small 4 |
|---|---|
| 20M in, 4M out, no caching | $16.00 / $60.00 / $5.40 |
| Same volume, 90 percent cached | $7.90 / $35.70 / $2.97 |
Medium 3.5 costs 3.75 times what the flagship costs on the same traffic, and 4.5 times once caching is on, because the gap widens as the cheaper input line does more of the work. Whatever reason there may be to run it, price is not one, and any pipeline that routes to Medium as a cost-saving step is doing the opposite of what it thinks. Our LLM cost calculator prices the same volume across every provider we track.
The catalogue is not all Mistral
One entry on that pricing page is worth reading carefully. GLM 5.3, at $1.40 input and $4.40 output, is served through Mistral but published by Z.ai, and Mistral labels it as third-party hosted. It is not a Mistral model, and comparison tables around the web routinely file it as one, which we did too before correcting it across this site.
Codestral, the code model, is priced at $0.30 and $0.90, and Leanstral 1.5 is currently free as a Labs experiment. The specialised models for documents and audio are billed per page or per minute rather than per token, so they sit outside this comparison entirely.
Where Mistral Large 3 sits in the wider market
At $0.50 and $1.50 it is among the cheapest flagship-tier models available anywhere, well below Claude Sonnet 5 at $2.00 and $10.00 and below the middle of the current GPT line. Its successor no longer undercuts it: the Mistral Large 4 launch pricing is over, and Mistral has since cut Large 4 for good to $0.70 in and $2.10 out, $0.07 for cached input, as it opened the weights, which is 1.4 times Large 3 on both lines (Mistral’s pricing page, read on 3 November 2026). Mistral deprecated Large 3 on 31 October 2026 and will retire it on 30 April 2027, pointing new integrations to Large 4. Only the open-weight hosts and the deep-discount providers go lower.
Price is not the whole decision, and this page does not pretend to rank models on anything else. But when a flagship costs less than most providers’ mid-tiers, it deserves a place in the evaluation rather than being filtered out on the assumption that flagship means expensive. Our full comparison of what every major model charges puts it against the rest of the market on a single reading date, and our Mistral API pricing page covers the rest of the catalogue.
Frequently asked questions
How much does Mistral Large 3 cost per million tokens?
$0.50 for input and $1.50 for output, with cached input discounted ninety percent. Read from the Mistral documentation on 16 August 2026.
Why is Mistral Large 3 cheaper than Mistral Medium 3.5?
Mistral does not explain it on the pricing page, and we are not going to invent a reason. What is verifiable is the grid: Large 3 is a third of Medium 3.5 on input and a fifth on output.
Is GLM 5.3 a Mistral model?
No. It is published by Z.ai and served through Mistral as a third-party hosted model, at $1.40 and $4.40 per million.
What is the cheapest Mistral model?
Ministral 3 3B at $0.10 for both input and output, followed by Ministral 3 8B and Mistral Small 4 at $0.15 input. Leanstral 1.5 is free while it remains a Labs experiment.
How does Mistral Large 3 compare with Claude or GPT on price?
It is substantially cheaper than either provider’s comparable tier, and its three-times output ratio is tighter than the five times both Anthropic and OpenAI charge. We put them side by side, read on the same day, in our cross-provider comparison.
What we verified and what we did not
Verified: input and output rates for Mistral Large 3, Medium 3.5, Small 4, the three Ministral 3 sizes, Codestral and the third-party GLM 5.3, plus the ninety percent cached input discount, all read from the Mistral pricing documentation on 16 August 2026, with the GLM row re-read on 3 October 2026, and with the Large 4 price and the Large 3 deprecation re-read on 3 November 2026.
Not verified, and therefore not asserted: why the tiers are priced in this order, which Mistral does not explain; the per-page and per-minute rates for the document and audio models, which are not comparable to token pricing and are not covered here; and any benchmark or quality claim, since this page prices models rather than ranking them.
Prices in this category move. Everything here carries the date it was read, and the grid is worth checking at the source before you commit a budget to it.