Mistral Large 4 vs GPT-6.1 Sol: One You Can Download, One You Can Only Rent

Almost every comparison of Mistral Large 4 and GPT-6.1 Sol on the open web is quoting numbers that stopped existing on 27 October. The pages say 38 and 52. Artificial Analysis rebuilt its Intelligence Index into the v5 series that day, re-ran the models, and now publishes 26 for Mistral Large 4 against 55 for GPT-6.1 Sol. That is not a change in either model — it is a different instrument, with nine evaluations instead of the old set and a much harder mix — but it does mean the gap you may have read about, fourteen points, is now twenty-nine on paper and not comparable to anything from before the rebuild. Comparisons across index runs are the most common error in this niche, and it is worth getting the current pair right before arguing about anything else.

What makes this matchup more interesting than a score difference is that the two products stopped being the same kind of thing this morning. GPT-6.1 Sol has been a rented model since OpenAI shipped it on 29 September 2026 at $2.00 and $10.00 per million tokens; you call it, you cannot keep it. Mistral Large 4, previewed on 6 October 2026, published its weights on 29 October 2026 at 14:51 UTC — 1,060.6 GB in FP8 and 610.1 GB in NVFP4, open to download, under a modified-MIT licence. One of these models has a meter that never stops. The other's meter ends the day you buy the hardware. That is the decision, and the benchmark table is downstream of it.

Two products, two dates, one of them finished

The asymmetry is not a quibble about maturity — it changes which questions are worth asking. GPT-6.1 Sol is a shipped, documented, independently measured model with ten API providers listed against it. Its rate card is settled, its 1M context is documented, and its position in OpenAI's line-up is below the Astra flagship rather than at the top. Mistral Large 4 spent twenty-three days as an API preview, and today it became a file. Its hosted rate card is settled at $0.70 input and $2.10 output per million tokens, and its future is now partly in the hands of whoever runs it rather than in Mistral's roadmap.

The six lines that matter

Same dimensions on both sides, current instrument, read from the two Artificial Analysis model pages and the vendors' own documentation on 29 October 2026:

• Intelligence Index — GPT-6.1 Sol 55 (v5 series, ranked seventh of 54 in its comparison class); Mistral Large 4 26 (ranked eighth of 19 in a different class — the two ranks are not on one scale).

• Cost per Index task — GPT-6.1 Sol $1.27; Mistral Large 4 $2.06.

• Token price — GPT-6.1 Sol $2.00 in / $10.00 out per million, with a $4.00 / $15.00 tier above 272,000 input tokens; Mistral Large 4 $0.70 in / $0.07 cached / $2.10 out.

• Output speed — GPT-6.1 Sol 79.1 tokens per second; Mistral Large 4 161.3.

• Context — 1M tokens on both, documented on both.

• Deployment — GPT-6.1 Sol: hosted only, no weights at any price. Mistral Large 4: hosted or self-hosted, 610.1 GB in NVFP4, eight Blackwell-class GPUs in the published vLLM recipe.

The Index is an average and averages hide the shape of a gap. On Artificial Analysis's individual rows the split is uneven rather than uniformly bad: HLE-Diamond 17% against 59%, Terminal-Bench-Science 0.1 at 4% against 58%, FrontierCode 2.0 at 15% against 41%, Terminal-Bench 4.0 at 25% against 56%, and GDPpdf at 18% against 31%. Mistral Large 4 is not behind by twenty-nine points on every task; it is behind by a lot on the evaluations that reward sustained multi-step reasoning, and roughly level on shorter single-pass work. If your workload looks like the second kind, the average is the wrong number to buy on. If it looks like the first, no discount closes it.

Why the numbers moved, and why that is not a scandal

The v5 rebuild re-ran the whole field on a new evaluation set, and roughly 60% of it is private — questions that were not publicly available before the rebuild. Models did not all move in the same direction: some dropped hard, some rose. Mistral Large 4 falling from 38 to 26 and GPT-6.1 Sol rising from 52 to 55 are two readings of the same event, and neither is a vendor having a bad quarter. The rule that follows is simple and easy to break by accident: never subtract a score from one index run against a score from another, and never quote a gap whose two halves were measured a week apart. The gap that exists today is twenty-nine points on one instrument.

Twenty-nine points is a lot, and it is not the whole story either. An Index of 26 describes a model that is competent at general work and unremarkable at the hard end; an Index of 55 describes one of the strongest models available. If your workload is agentic coding over a large repository, the answer is already decided and no pricing argument changes it. The rest of this piece is about the workloads where it is not decided.

The reversal survived the rebase

The one finding from the October preview period that held up under the new instrument is the price-per-task inversion, and it held up with fresh numbers. 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 62% more to complete the same evaluation task — $2.06 against $1.27. That is not a pricing trick. A reasoning model's bill is dominated by how many tokens it chooses to emit, and Mistral Large 4 emits a great many: Artificial Analysis records 110 million output tokens across its verbosity sample against a 100 million median, while GPT-6.1 Sol sits at 64 million, below the median. Cheaper tokens multiplied by more tokens is not a saving, and the invoice will say so.

The practical consequence is narrower than "the cheap model is secretly expensive". For short-output work — classification, extraction, routing, anything that ends after two hundred tokens and never enters a reasoning loop — the token price is the bill, and $0.70 against $2.00 is the whole argument. The inversion bites on long-horizon reasoning, which is exactly where the twenty-nine-point gap also lives. Two different measurements pointing the same direction is a reason to trust the direction, not a reason to generalise it to every workload.

Speed runs the other way, and so does patience

Mistral Large 4 is more than twice as fast in output: 161.3 tokens per second against 79.1. It also returns a first token in 1.42 seconds, slightly better than the 1.68-second median for its comparison set. GPT-6.1 Sol's published figure for time to first answer token is 249.48 seconds, against a 3.53-second median, which is the kind of number that needs explaining rather than quoting.

That 249 seconds is almost certainly not latency in the network sense. It is the time the model spends reasoning before it emits the first answer token, measured at maximum reasoning effort, and it is the honest way to show what a long reasoning budget costs a user who is waiting. Our own routing data makes the distinction concrete: across 584.7 million GPT-6.1 Sol tokens carried through OrcaRouter in the seven days to 29 October, the p50 time to first streamed token was 1.36 seconds while p95 was 10 seconds. The first token arrives quickly; it is the first answer that takes minutes, and the flags in the API decide which one your user experiences. If your application streams reasoning tokens, the wait is populated and tolerable. If it does not — if a user stares at a spinner until the first answer arrives — then a four-minute first token is a product constraint that no benchmark score compensates for, and a model that answers in 1.4 seconds will feel better than one that is twenty-nine Index points stronger. This is the axis on which the cheaper, lower-scoring model most often wins real deployments, and it does not appear in any comparison table.

What changed this morning: the meter ends

Until today the self-hosting argument for Mistral Large 4 was a promise, and promises do not survive procurement review. It is now a repository. Two checkpoints, one licence file, and a vendor-published serving recipe: vLLM with tensor parallelism 8 and an FP8 MLA KV cache, SGLang at tp-size 4, both with an embedded DSpark draft head for speculative decoding rather than a second model.

What that buys is not a cheaper token — it is the end of the token as the unit of cost. On owned hardware the bill becomes depreciation, power, rack space and the engineer who owns the cluster, and a workload that runs at high, steady volume converts from a variable cost into a fixed one. That is the argument for a European bank with a data-residency rule, a defence contractor whose procurement forbids foreign inference, or a manufacturer embedding a model in a product line for five years. It is not an argument for a startup with spiky traffic and no operations team: 610 GB of weights, eight Blackwell GPUs and a 1M-token KV cache is a capital project, and the hosted $0.70 rate is cheaper than the capital for a very long time.

The licence is the part people skip. It is a modified MIT with one clause that matters: running the weights as a model-as-a-service offering requires a separate commercial licence from Mistral. Read the definition carefully, though — it excludes embedded end-user products and "mere relaying of requests to models hosted by others", so a gateway passing traffic to Mistral-hosted inference is carved out while a company standing up its own inference farm and reselling access is not. If your plan is "download it and become a provider", read the clause before the hardware quote.

Where each one is actually the right answer

• Agentic coding, long-horizon tool use, large-repository work — GPT-6.1 Sol. The Index gap is widest exactly here and the price premium is the cost of being right.

• High-volume, short-output extraction and classification — Mistral Large 4 at $0.70 input. The reasoning ceiling never engages, so the Index gap never applies.

• Regulated, jurisdiction-bound deployment — Mistral Large 4, and not because it is better. GPT-6.1 Sol cannot be self-hosted at any price, which ends the comparison.

• Interactive assistants where the first token is the product — Mistral Large 4 on speed, unless you stream reasoning and can absorb the wait.

• Defensive security tooling — Mistral Large 4, on the strength of a vendor-reported 96.6% refusal rate on malicious cyber requests and a cyber benchmark set that Mistral organised its launch around. Those figures are Mistral's, not an independent lab's, and should be treated as claims until someone reruns them.

• Anything where the answer must be right and the latency budget is generous — GPT-6.1 Sol. Twenty-nine points is not a rounding error.

If you are going to run both

Mixed estates are the normal outcome of a comparison like this, and the operational question becomes how two very different models sit behind one bill. GPT-6.1 Sol is on OrcaRouter at exactly OpenAI's list price — $2.00 and $10.00 per million tokens, with the above-272K tier at $4.00 and $15.00 — routed at 0% markup, so the rate you see is the rate OpenAI charges. Mistral Large 4 is not in our catalogue: 214 models, no Mistral endpoint, and if you want to call it you go to Mistral's own API or to your own vLLM deployment. What the platform does carry, for the hosted half, is the routing DSL — point a workload at one model and move it without a code change — and automatic failover, which is what makes trying a new model on a production path survivable. Wire the self-hosted endpoint in behind the same pattern and the two halves of this comparison stop being a decision you have to make once.

Two questions, not one

The comparison pages will tell you GPT-6.1 Sol wins, and on the only independent instrument both models share, it does — 55 against 26, at 62% less per completed task and a little over half the output speed. That is the correct answer to "which model is better", and if that were the only question the market would have settled this in September.

The question that opened this morning is different: which model can you still be running in three years, on hardware you control, under a licence you have read. For a great many teams the answer will remain "neither, we rent" — and then GPT-6.1 Sol is the better rent. For the ones whose constraint is ownership rather than capability, the model that scores 26 stopped being a compromise today and became the only option on the table. The two answers can both be right, which is why the interesting work is not picking a winner but deciding which question your organisation is actually asking.