Mistral Large 4 launched at half price and Mistral hasn't said when the sale ends. At $0.68 and $2.09 it costs what DeepSeek V4-Pro costs off-peak. At the $1.36 and $4.18 list price, it's GLM-5.3 money.

Short version: today Large 4 costs about what DeepSeek V4-Pro costs off-peak, and it scores about the same on Artificial Analysis. When the sale ends it costs twice that. GPT-6 Luna gets the same index score for roughly an eighth of the money per task. Pick Large 4 for the things only it does: EU hosting, weights you can download at the end of the month, and security work the US frontier models refuse.

The price on the page, and the one crossed out

Large 4 went live on October 6 as a public preview, model ID mistral-large-4 (or the pinned mistral-large-4-0). The announcement's spec card says $1.36 in and $4.18 out. Open the model page in Mistral's docs and those figures are struck through, with $0.68 and $2.09 next to them. That is exactly half. OpenRouter's endpoint data marks the route with a 0.5 discount too.

Mistral docs, model page and pricing table, read October 10, 2026. Batch rows are input / output. Batch cached input is $0.035 on sale and $0.07 at list. The batch tab only renders in a browser.

There's no end date. We checked both docs pages for one and found nothing, and the launch post doesn't mention a sale at all. We have written about discounts with published expiry dates. This one is worse, because you can't put it in a calendar.

Against Large 3 ($0.50 / $1.50), the sale price is 36% more on input and 39% more on output. At list it's 2.7x and 2.8x. Large 3 isn't going anywhere yet, so the cheaper option is still there.

One month, 50M tokens in and 10M out

A mid-sized agent or RAG workload with no caching. Every model here has a published per-token price we could check.

Our arithmetic at standard list or sale rates, no cache, no batch. Luna assumes prompts under its 272K step-up. DeepSeek's peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays.

The sale puts Large 4 $2.10 a month above V4-Pro off-peak. The list price puts it $4.20 under GLM-5.3. Same model in both rows, and the only thing that changes is whether Mistral keeps the discount.

It talks a lot

Per-token prices only tell half of it. Artificial Analysis measured Large 4 Preview (reasoning) at 38 on its Intelligence Index, and it took 200M output tokens to run the suite. The median model needs 82M. So what one task really costs looks like this:

Cost per Intelligence Index task from Artificial Analysis, read October 10, 2026. AA prices Large 4 at the $1.36 / $4.18 list rate. The sale row is that figure halved, our arithmetic.

Even on sale, Large 4 spends about eight times what Luna spends for the same 38. V4.1 Flash scores a point higher for under half the cost. It's a preview build, so this could change. We would still test with your own prompts before believing the per-token sticker.

AA hasn't published speed or time-to-first-token for it yet. You can price your own token counts in the calculator, which now has Large 4 at the sale rate.

Where Mistral says it wins

Mistral picked its comparisons mostly from the open-weight pack and from security work. All of these are vendor-reported.

Mistral's Large 4 announcement, October 6, 2026.

The vulnerability number is the interesting one. A model that will reproduce and patch an exploit when Opus 5.5 and Astra decline is a product in itself, whatever it costs per token. If you work in that niche, you aren't really comparing prices here. Everyone else should be.

1M tokens, unless you ask for Europe

Mistral's docs say 1M context. Artificial Analysis says 524K. Both are right. OpenRouter lists 1,048,576 on the global route and 524,288 on the EU one, so the regional deployment gets half the window. The regional endpoints, EU or US, also bill 1.1x and don't take batch jobs.

That matters because EU residency is half the reason to pick Mistral. On the sovereign route you pay $0.748 and $2.299 at the sale price, and you get half the context. Max output is 262,144 tokens on every route, according to OpenRouter. Mistral doesn't state one.

Weights at the end of the month

Mistral calls Large 4 open-weight, and the Hugging Face repo exists, but it's a placeholder with an ETA of October 31. No license is named yet. Large 3 shipped under Apache 2.0 and Medium 3.5 under a modified MIT, so we'd guess one of those, but it is a guess.

For now Mistral is the only provider. We found no Bedrock, Vertex or Foundry listing. Once the weights are public, other hosts can sell it, and with roughly 1T total and 49B active parameters (52B with embeddings) it's a big model to run, so we wouldn't expect them to beat the sale price by much. They may well undercut the list price. If you're planning beyond a few weeks, that is the number to watch.

Worth it at $1.36?

If you need a European provider, an open-weight model you can later host yourself, or a model that will do offensive security work, Large 4 is the newest serious option and the sale makes it cheap to try. Do your evaluation now, while it's half price.

If you just want the most intelligence per dollar, it isn't the pick. Luna, V4.1 Flash and MiMo-V2.6-Pro all score the same or higher for less per task. And if you build a budget on $0.68, write $1.36 next to it. You can line the models up on the pricing table.

Where the numbers come from

• Mistral: Large 4 announcement - release, list price, parameters, benchmarks, weights timing
• Mistral docs: Large 4 model page and pricing table - sale and original prices, model IDs, preview status
• Mistral docs: regional inference - 1.1x regional pricing, no batch
• Hugging Face: Mistral-Large-4-1T-A52B - weights placeholder, ETA October 31
• OpenRouter: Mistral Large 4 - endpoint context lengths, max output, discount flag
• Artificial Analysis: Mistral Large 4 - Intelligence Index, tokens used, cost per task
• DeepSeek: pricing - V4-Pro and V4.1 Flash, peak and off-peak