Mistral Large 4 opened its weights and cut its API price to $0.70/$2.10. It now beats GLM-5.3 per task, but GPT-6.1 Terra scores 16 points higher for an eighth of the cost.
A week after its launch sale ran out, Mistral put Large 4 back on sale for good. On Thursday it posted the weights and a new API price in the same announcement. The cut is real. Whether it makes Large 4 a good buy depends on who you compare it with.
Image source: Mistral AI
Old price, new price
The announcement says the new rates apply "starting today", October 29, and the model page matched them when we checked this morning. Mistral never prints the old price next to the new one, so the "before" row is what that same docs page showed when we last read it on October 25.
| Mistral Large 4 | Input | Cached | Output |
|---|---|---|---|
| Launch sale, Oct 6 to ~Oct 20 | $0.68 | $0.07 | $2.09 |
| List, read Oct 25 | $1.36 | $0.14 | $4.18 |
| From Oct 29 | $0.70 | $0.07 | $2.10 |
| Change vs Oct 25 | -48.5% | -50% | -49.8% |
Per million tokens, Mistral's global API. The EU route on OpenRouter bills 1.1x ($0.77 / $0.077 / $2.31) with a 512K window.
That puts the permanent price a cent or two above the two-week sale. Nothing on the page puts an end date on it. Two gaps remain: Mistral's general pricing page still doesn't list Large 4, and nobody has published a Large 4 batch rate. The site-wide rule is 50% off for batch, which would make it $0.35 / $1.05, but that's our arithmetic, not a quoted price.
Per task, it beats GLM-5.3 and little else
Artificial Analysis already re-priced its Large 4 runs at the new card. Here is what one task on its Intelligence Index v5.0 costs, cheapest first. The number before each name is the score.
- 29.6 Claude Haiku 5.5, high$0.16
- 42.3 GPT-6.1 Terra, high$0.26
- 32.5 DeepSeek V4.1 Flash, max$0.60
- 23.1 Large 4, medium$1.41
- 23.1 GLM-5.3, low$1.57
- 26.0 Large 4, high$2.06
- 46.0 Kimi K3.1 Pro$3.68
- 31.7 GLM-5.3, max$5.01
Artificial Analysis Intelligence Index v5.0, read November 1, 2026, at each provider's list price. v5.0 scores are not comparable with the v4 score of 38 in our launch post.
Against the model Mistral keeps putting next to it in its own charts, Large 4 now wins. Large 4 at medium and GLM-5.3 at low land on the same 23.1, and Large 4 gets there 10% cheaper. Push both to the top and GLM-5.3 max scores 5.7 points more but costs $5.01 a task to Large 4 high's $2.06.
Step outside open-weight models and it looks much worse. GPT-6.1 Terra at high scores 42.3 for $0.26, 16 points more for an eighth of the money. Claude Haiku 5.5 at high beats it by 3.6 points for $0.16. Before Thursday Large 4 high would have cost about $4.10 a task, so the cut halved the gap. It didn't close it.
One more thing before you trust the launch charts. Mistral reports 41% on Terminal-Bench 4.0. Artificial Analysis ran it and got 24.7%. On FrontierCode 2.0 it's 19.7% from Mistral and 15.0% from AA. Different harnesses explain some of that, but we'd plan around the lower numbers.
A month of agent traffic
The workload we use across these posts: 1 billion input tokens, 85% served from cache, and 50 million output. Cache writes are billed where the provider charges for them (OpenAI 1.25x input, Anthropic and Moonshot at their 5-minute write rate). Mistral and Z.ai list no write fee.
| Model | 150M uncached | 850M cached | 50M out | Month |
|---|---|---|---|---|
| Claude Haiku 5.5 | $18.75 | $8.50 | $25.00 | $52.25 |
| DeepSeek V4.1 Flash, peak | $45.00 | $5.10 | $60.00 | $110.10 |
| Mistral Large 4, now | $105.00 | $59.50 | $105.00 | $269.50 |
| GPT-6.1 Terra | $187.50 | $42.50 | $250.00 | $480.00 |
| Mistral Large 4, Oct 25 | $204.00 | $119.00 | $209.00 | $532.00 |
| GLM-5.3 | $210.00 | $221.00 | $220.00 | $651.00 |
| Kimi K3.1 Pro | $300.00 | $170.00 | $500.00 | $970.00 |
List prices read November 1, 2026. Haiku 5.5 at its under-100K prompt rate; DeepSeek at peak hours (off-peak halves it). No batch.
On equal token counts, Large 4 is now cheaper than Terra and GLM-5.3 and saves $262.50 a month against its own October price. But token counts are not equal. Large 4 high wrote 110 million output tokens on the AA index, and that is why it costs eight times Terra per task while looking cheaper here. Measure your own output before picking from this table.
The weights, and what running them costs
Two builds sit on Hugging Face. The FP8 one is about 1,061 GB and Mistral's serving example spreads it over eight GPUs. The NVFP4 one is about 610 GB, targets B200, B300 and GB200/GB300, and its SGLang example runs on four (the vLLM one uses eight). Mistral's docs give 1.05 trillion parameters in total; the model card says 49 billion are active, 52 billion counting embeddings, with a 1M context window.
Pick your own GPU rate. At $5 per GPU-hour, which is our assumption and not a quote, four cards running all month cost $14,600. That is 54 times our agent month on Mistral's API. You'd need to push around 54 billion input tokens and 2.7 billion output tokens through those four GPUs every month just to break even, and we haven't measured whether four cards can serve that.
So at this price, self-hosting is about control, not savings: EU or on-prem data rules, fine-tuning, or a product that can't send prompts out. The license covers that. What it blocks is reselling. Under the Mistral Ecosystem Open-Weight license, anyone running a Model-as-a-Service business needs a separate commercial license, which is probably why Mistral is still the only paid API on OpenRouter three days in. Don't expect a cheaper Together or Fireworks endpoint to undercut $0.70 soon.
Where Large 4 fits now
If you run GLM-5.3 and like it, try Large 4 at medium. At equal scores it's 10% cheaper per task and half the price per token, and it adds EU hosting. If you already ran Large 4 during the sale, your bill goes back to almost exactly what it was.
Terra or Haiku 5.5 win on the per-task numbers if you don't need open weights or a European provider. Kimi K3.1 Pro's weights are due November 4 and it scores 20 points above Large 4, so anyone shopping for a model to host themselves should wait three days before committing.
We've updated Large 4 on its model page. To see your own month, put your token counts into the calculator or check GLM-5.3's page side by side .
Sources
- Mistral: Mistral Large 4 open-weight release - new price, license, Mistral-reported benchmarks, hero image
- Mistral docs: mistral-large-4-0 - current rates, 1M context, parameter counts, preview status
- Hugging Face: Mistral-Large-4-1T-A52B and the NVFP4 build - file sizes, serving examples, license text
- Artificial Analysis: Mistral Large 4 and leaderboard - Index v5.0 scores, cost per task, output tokens, Terminal-Bench 4.0
- The Next Web - four B200 or B300 GPUs for the quantised build
- Z.ai, DeepSeek, Kimi, OpenAI and Anthropic pricing pages - comparison rates