GPT-6.1 Sol vs Mistral Large 4: The Cheaper Model Per Token Is the More Expensive Model Per Task

The interesting thing about putting GPT-6.1 Sol next to Mistral Large 4 is not which one is smarter — on the only independent measurement both have, that is not close. It is that the model with the lower sticker price is the one that costs more to finish a job with. Mistral Large 4, previewed by Mistral on 2026-10-06, is priced at $0.68 per million input tokens and $2.09 per million output while the preview lasts, against GPT-6.1 Sol's $2.00 and $10.00 since Open​AI released it on 2026-09-29. And yet Artificial Analysis's cost-to-run figures put Mistral Large 4 at $1.13 per task against GPT-6.1 Sol's $0.72 — because the cheaper tokens are not the whole bill, and because reasoning models spend tokens in amounts you do not choose from a price list.

Everything below comes from Mistral's own announcement and documentation, Open​AI's model and pricing pages, Artificial Analysis runs for each model, and our own serving data, read on 2026-10-12. Where a number is the vendor's own and nobody has rerun it, the sentence carrying it says so.

Two dates, and only one of them is finished

Before any spec, the state of the two products is different, and that difference does more work in this comparison than any benchmark cell.

• GPT-6.1 Sol is a shipped model. It went live on the Open​AI API on 2026-09-29 at $2.00 / $10.00 per million tokens, with $0.10 cached input and $2.50 cache writes, and it is the model in the API's default slot. Its rate card is published, its context window is documented, and independent instruments have run it. What you read about it you can call this afternoon.

• Mistral Large 4 is a preview. Mistral announced it on 2026-10-06 as a 1.05-trillion-parameter mixture-of-experts model with 52 billion parameters active per token, available on the vendor's own API at preview pricing. The open weights are promised before the end of October, which is why this is a preview and not a finished release. Nothing about its future list price, future context window or future licence is settled by the preview.

That asymmetry sets the shape of everything below. When one side of a comparison is a product and the other is an announcement with a date attached, most of the interesting questions are about timing rather than capability — and the one question people actually ask ("should I switch?") has different answers in October and in November.

The one number both sides agree on, and the 14 points behind it

Artificial Analysis has an independent Intelligence Index row for both models, and there is no ambiguity in it:

• Intelligence Index — GPT-6.1 Sol 52 at maximum reasoning effort; Mistral Large 4 38.

• Cost to run the Index — GPT-6.1 Sol $0.72 per task; Mistral Large 4 $1.13 per task.

• Output speed — GPT-6.1 Sol 65.4 to 66.8 tokens per second across runs; Mistral Large 4 is not the fast side of this pairing either.

Read those three lines together and the shape of the trade appears. Mistral Large 4 charges roughly a third of GPT-6.1 Sol's input rate and a fifth of its output rate, and still costs about 57% more to complete the same evaluation task. That is not a pricing trick or a measurement artefact; it is what happens when a model reaches a given quality level by spending more tokens to get there. A reasoning model's bill is dominated by how many tokens it decides to emit, not by the per-token rate on the card, and a 14-point Index difference is a large amount of extra thinking to buy back.

None of which makes Mistral Large 4 a bad model. An Index of 38 is a competent general-purpose frontier-adjacent score, and for a great many production tasks the extra 14 points buy nothing you can use. The point is narrower and more practical: do not buy this model on the assumption that the cheaper card makes the month cheaper. Measure cost per finished unit of work, which is the number Artificial Analysis is actually reporting and the number your invoice will reflect.

Where Mistral's case is genuinely strong, and it is not the price

Three arguments for Mistral Large 4 survive contact with the Index number, and two of them are not about the model at all.

• European training and sovereign hosting. Mistral is the European frontier lab, and the announcement sits under a sovereign-cloud message rather than a benchmark message. For a European public-sector buyer, a regulated bank, or anyone whose procurement rules name a jurisdiction, that is a requirement rather than a preference — and it is the one axis on which GPT-6.1 Sol cannot compete at any price. If your constraint is "the weights and the serving must be European," the comparison ends here and the Index number is irrelevant.

• Open weights are coming, and self-hosting changes the arithmetic. Mistral has committed to releasing the weights before the end of October. A 1.05T/52B mixture-of-experts is not a model most teams will run on their own hardware — 52B active parameters is the friendly half of that sentence, the trillion-parameter total is the unfriendly half — but for organisations that already run that class of infrastructure, or that have a data-residency rule no hosted API can satisfy, open weights turn a metered API into a fixed cost. GPT-6.1 Sol will never be self-hostable.

• The preview price is real for as long as it lasts. At $0.68 / $2.09 the preview discount is roughly half the announced list rate of $1.36 / $4.18. If a workload is cheap enough that token count is genuinely the binding constraint — bulk classification, extraction, routing, anything with short outputs and no reasoning loop — that price is the best on this page, and the Index gap does not apply to work that never needed an Index of 52. Just plan the switch when the preview ends rather than being surprised by it.

What is not an argument is raw capability at the top end. On agentic coding and long-horizon tool use, the gap is visible and it is the direction you would expect from the Index: GPT-6.1 Sol's Artificial Analysis Terminal-Bench 4.0 row reads 56.06, and the same instrument put its predecessor, GPT-6 Sol, at 43.94 — a roughly 12-point generational move on a benchmark that is specifically about finishing tasks in a shell.

The context number that disagrees with itself

This is the one specification on the page you cannot quote confidently, and it is worth flagging rather than smoothing over.

Mistral's own documentation for Mistral Large 4 describes a context window of up to 1 million tokens. The independent record for the model at Artificial Analysis carries 524,000. Those two numbers are not a rounding difference; they are a factor of two, and they point at different things. The most likely explanation is that one describes the architecture's ceiling and the other describes what is actually being served during a preview — a distinction that matters enormously if you were planning to load a large repository into a single call. We have not seen either figure reconciled by the vendor, and until it is, treat the smaller number as the one you can build against.

By contrast, none of GPT-6.1 Sol's headline limits have this problem. Its context window is 1,050,000 tokens, its maximum output is 128,000 tokens, and its knowledge cutoff is 2026-04-30 — all documented on the vendor's model page and consistent with what is served through the API.

What a thousand agent tasks cost on each

Take a workload that is ordinary for either model: one call that carries 40,000 input tokens and produces a 2,000-token answer, run a thousand times. To keep the two rate cards comparable, assume the input goes in fresh every time rather than being served from a cache.

• On GPT-6.1 Sol — 40,000 input tokens at $2.00 per million is $0.080, and 2,000 output tokens at $10.00 per million is $0.020. That is $0.100 a call, or about $100 for the thousand.

• On Mistral Large 4 at preview rates — 40,000 input tokens at $0.68 per million is $0.027, and 2,000 output tokens at $2.09 per million is $0.004. That is $0.031 a call, or about $31 for the thousand.

So the price gap is real and it is roughly three to one on a shape where output is short — with one honest caveat that cuts the other way. GPT-6.1 Sol charges $0.10 per million for a cached prefix, and Mistral's preview documentation does not establish an equivalent discount. Run the same loop with the 40,000 tokens served from cache and the GPT-6.1 Sol side falls to about $0.024 a call, which is below the Mistral figure. The rate cards decide this comparison only once you know whether your input repeats. Now change the other parameter and watch the second spread open: run the same loop at the reasoning-heavy end, where Mistral Large 4 is doing the token spending its Index of 38 implies, and the $1.13-per-task figure on a comparable evaluation says the Mistral side of the invoice grows faster than the rate card suggests. The honest summary is that Mistral Large 4 wins the arithmetic on short-output work at fresh input rates and loses it on cached and long-horizon agentic work, and which one you are running is a question about your workload, not about the models.

Both of these are exactly the kind of comparison a router is for. GPT-6.1 Sol is in our catalogue at Open​AI's list rates with 0% markup — what Open​AI publishes is what a caller pays, and when a vendor changes a price the catalogue row changes with it rather than waiting for a reseller to notice. That means the per-token arithmetic above can be checked against the live row rather than against this page. Mistral Large 4 is not something we serve, so if you want it, the vendor's own API and several third-party platforms are where it lives.

What to do on Monday

There is no version of this where the two models are substitutes for the same job today, and the reason is timing rather than quality.

If you need a frontier agentic coding model this week — long tool loops, repository-scale context, computer use, document-heavy professional work — the answer is GPT-6.1 Sol, it is the only one of the two that is a finished product, and its price is published and stable. Budget for the fact that its output meter is where the money goes, and keep the working set under the 272,000-token threshold above which the whole request reprices.

If you are in Europe and your requirements name a jurisdiction, or if you intend to self-host because no API satisfies your data rules, Mistral Large 4 is the only candidate here regardless of scores — but the part of it you are waiting for, the weights, does not exist yet, and the context figure you would size your infrastructure around is currently disputed by a factor of two. Watch for the weights release and for the vendor to reconcile 1M against 524,000 before you commit a deployment to it.

And if what you actually want is to use both — Mistral Large 4 for the cheap high-volume tier and GPT-6.1 Sol for the hard tail — the practical constraint is that the two live on different keys, different billing and different failure modes, which is the problem a single endpoint with automatic failover across 200-plus models exists to solve. In the meantime the row worth opening is the one that actually exists: the GPT-6.1 Sol model page at https://www.orcarouter.ai/models/openai/gpt-6.1-sol.