Mistral Large 4 Pricing vs Claude and GPT: What a Task Costs

Mistral Large 4 went into public preview on October 6 at a per-token price well under Claude Opus 5.5 and GPT-6 Astra. Open weights are promised by the end of the month. If you run an LLM feature with a real monthly bill, the obvious question is whether moving it would save money. The price sheet answers that badly: the launch discount isn’t the price, cached input narrows the gap, and the bill depends on tasks finished rather than tokens sent.

Mistral Large 4 pricing next to Claude and GPT

Here are the list prices per million tokens, as published on each provider’s pricing page on October 9, 2026. Mistral’s figures come from its model page (opens in a new tab), which shows a discounted launch price next to the standard one.

Claude prices are from Anthropic’s pricing page (opens in a new tab) and GPT prices from OpenAI’s (opens in a new tab). OpenAI’s figures are for prompts up to 272K input tokens. Above that, input doubles and output rises by half.

At standard prices, Mistral Large 4 is about 2.9x cheaper than Opus 5.5 on input and 4.8x cheaper on output. Against Sonnet 5.5 and GPT-6.1 Sol, which share the same $2/$10 price, the gap shrinks to about 1.5x on input and 2.4x on output. Those two models are the honest comparison for most production workloads, because few teams run every request through a flagship.

Budget at the standard price, not the launch price

The launch price halves everything, and it’s the number that gets quoted. Don’t put it in a forecast. Mistral’s model page shows it as a reduction from the standard price and gives no end date. The announcement doesn’t mention it at all. The model is also a preview. Previews can change behaviour, model IDs and limits before general availability, and the regression suite you’d need to trust it has to be run again when they do.

Build the business case on $1.36 and $4.18. If the move only pays off at the launch price, it doesn’t pay off.

Cached input is where the gap narrows

The table hides the line that matters most for agents and long-context features. Mistral Large 4’s standard cached-input price is $0.14 per million tokens. Sonnet 5.5 and GPT-6.1 Sol charge $0.10. On the tokens that dominate an agent loop, the re-sent conversation history, the cheaper model is the more expensive one.

Take an illustrative agent run of 200,000 input tokens, 180,000 of them cache hits, with 8,000 output tokens. Cache writes are left out to keep the arithmetic readable.

Mistral Large 4 still comes out cheaper, at about 62% of the Sonnet 5.5 cost for this run. That’s a 38% saving, not the 2.4x that the output column suggests. The more cache-heavy your traffic, the smaller the saving. A workload that is mostly short, uncached prompts with long answers sits closer to the headline ratio. Model your own mix of uncached input, cached input and output before you believe either number. Your current provider’s usage records already split them.

Cost per finished task decides the switch

A provider bills tokens per attempt, but you pay for finished tasks. Three things move the cost per task without touching the price per token.

Attempts per success. If a cheaper model fails a task that the current one finishes, you pay for the failed attempt, the retry and often a fallback call to the expensive model. In the run above, Mistral Large 4 costs $0.086 against Sonnet 5.5’s $0.138. If it needs more than about 1.6 attempts for every task Sonnet finishes in one, the saving is gone. That break-even is easy to cross on hard tasks. In Mistral’s own blind human evaluation of coding quality, run with Surge AI, Mistral Large 4 scored 3.74 out of 5 against Claude Opus 5’s 4.22 (opens in a new tab). That’s a vendor-run evaluation against Anthropic’s previous flagship, so treat it as a hint about where the gap sits rather than a verdict on your workload.

Token counts. The same prompt isn’t the same number of tokens on every model. Anthropic’s pricing page notes that Claude 4.7 and later use a newer tokenizer that produces about 30% more tokens for the same text than earlier Claude models. Mistral’s tokenizer will differ again. Per-million prices only compare cleanly once you’ve counted the tokens your prompts actually produce on each model.

Reasoning tokens. Both Claude and OpenAI bill hidden reasoning as output tokens. OpenAI’s reasoning guide says a request can use anywhere from a few hundred to tens of thousands of them. Output is the most expensive column for every model in the table. A model that thinks twice as long on your prompts can erase much of a price advantage. Claude reports the reasoning share in usage.output_tokens_details.thinking_tokens and OpenAI in output_tokens_details. Compare that field, not just the final answer length.

Run the comparison on your own workload

The comparison worth trusting is a small one on your own traffic. If you already have an eval, this is an afternoon’s work. If you don’t, a 20-task eval on your own repository shows the shape for coding work, and the same approach works for any task with a checkable result.

1. Pick one task type with a clear pass or fail, and sample 50 to 100 real requests from it.
2. Run them on the current model and on Mistral Large 4, with the same prompts, the same tools and a fixed reasoning effort.
3. Record input, cached input, output and reasoning tokens per request from each provider’s usage fields, and price them at standard rates.
4. Divide the total cost by the number of tasks that passed, not the number attempted.
5. Run it again on a second day. One green run of a non-deterministic system tells you little.

Step 4 is the one that changes decisions. Before switching anything in production, pin the current prompts to a regression suite, because the prompts you tuned for one model rarely carry over unchanged.

What open weights change, and what they don’t

Mistral’s Hugging Face page lists the weights as an upcoming release with an ETA of October 31, 2026. It describes a 1-trillion-parameter model with 49 billion active parameters per token, or 52 billion counting embeddings and output layers. The license hasn’t been published.

The low active-parameter count makes each token cheap to compute, but every parameter still has to sit in memory. At 8 bits per parameter, the weights alone come to roughly a terabyte before any KV cache. That needs a multi-GPU cluster, which very few teams will run to save on an API bill. Open weights matter in other ways. Third-party hosts can compete on price, which caps what Mistral can charge. And regulated workloads can run the model where their data has to stay. If you do host it, the defaults on a self-hosted LLM server need locking down before anything else.

When to move a workload

Move a workload to Mistral Large 4 when three things are true. It’s mostly short or uncached prompts, where the price gap is widest. It passes your own eval at a success rate close to the current model’s, so the cost per finished task drops at standard prices. And you can afford to rerun that eval when the preview changes.

Leave cache-heavy agent loops and the hardest tasks where they are for now. The saving there is thinner than the table suggests, and the only quality comparison published so far, Mistral’s own, puts the model behind Claude Opus 5 on coding. If you already route easy requests to a smaller model, a model cascade is the natural place to add Mistral Large 4: as a middle tier that one task type escalates to, measured against the tier it replaces.

Sources

• Mistral Large 4 model page (Mistral docs) (opens in a new tab)
• Mistral Large 4 announcement (opens in a new tab)
• Mistral-Large-4-1T-A52B (Mistral on Hugging Face) (opens in a new tab)
• Claude pricing (opens in a new tab)
• Claude extended thinking (opens in a new tab)
• OpenAI API pricing (opens in a new tab)
• OpenAI reasoning models guide (opens in a new tab)