GPT-6 Astra vs Mistral Large 4: A Capability Gap, a Price Gap, and Three Things Benchmarks Miss

Mistral Large 4 is roughly three to seven times cheaper than GPT-6 Astra and roughly sixteen Intelligence Index points behind it. On the independent numbers read in the week of 8 October 2026, Mistral Large 4 Preview sits at 38.4 on Artificial Analysis' Intelligence Index v4.3 and GPT-6 Astra at 54.7 — a gap wider than the one between Astra and any other model in this bracket — while the rate cards run $0.68 and $2.09 per million tokens against $10.00 and $50.00. If capability per dollar were the only question, the answer would be Astra for hard work and Mistral Large 4 for volume, and the article would end here.

It does not end there, because three of the differences that decide this matchup do not appear on either vendor's benchmark table. Mistral Large 4 arrived on 6 October 2026 as a public preview whose weights have not been released, served from Mistral's own European datacentres, and marketed on a refusal behaviour that closed models cannot replicate at any price. GPT-6 Astra shipped on 3 September 2026 as the top of OpenAI's line, is closed, moderated, and in this comparison is the model that says no. That is a different axis from the one the index measures.

The specification, both sides on every line

• Release — GPT-6 Astra 3 September 2026, generally available vs Mistral Large 4 6 October 2026, public preview

• Parameters — Astra undisclosed vs Mistral Large 4 1.05 trillion total with 49 billion active per token, plus a 1.6-billion-parameter vision encoder

• Context window — Astra 1,050,000 tokens vs Mistral Large 4 advertised at 1M but read at 524,288 by Artificial Analysis on the configuration it is testing

• Maximum output — Astra 128,000 tokens vs Mistral Large 4 not yet documented

• Input price — Astra $10.00 per 1M vs Mistral Large 4 $1.36 on the published card, currently charged at a promotional $0.68

• Output price — Astra $50.00 per 1M vs Mistral Large 4 $4.18 on the card, currently $2.09

• Cached input — Astra $1.00 per 1M vs Mistral Large 4 $0.14, currently $0.07

• Long-context tier — Astra reprices the whole request to $20.00 and $75.00 above 272,000 input tokens; Mistral publishes no comparable step

• Weights — Astra closed vs Mistral Large 4 promised under an unnamed open licence by the end of October 2026, not yet released

• Where it runs — Astra on OpenAI's infrastructure, Azure and AWS Bedrock vs Mistral Large 4 on Mistral's own European cluster today, self-hostable once the weights land

• Independent Intelligence Index — Astra 54.7 vs Mistral Large 4 Preview 38.4

• Independent long-context recall — Astra 80.7 vs Mistral Large 4 81.3, the one capability line Mistral leads

The index gap is real, and it is not the whole model

Fifty-four against thirty-eight is not a rounding difference. On the same evaluator's Terminal-Bench 4.0 harness Astra records 59.1% against Mistral Large 4 Preview's 26.8%, and on Artificial Analysis' own page Mistral Large 4 is flagged as averaging roughly 507 seconds per task where Astra's index runs are priced at about $3.26 of compute each. A model that takes eight minutes per task and scores sixteen points lower is not competing in the same workload class on agentic coding.

Two lines run the other way, and both are worth carrying into any decision. Mistral Large 4 leads on long-context recall, 81.3 against 80.7, which is a genuine strength for the document-heavy and retrieval-heavy work where a European buyer is most likely to want it. And it is a preview of a 1.05-trillion-parameter mixture-of-experts that Mistral says is still in an unsaturated reinforcement-learning run — the vendor's own words are that it expects "large and rapid improvements in the weeks and months to come." The model you benchmark today is the worst version of it you will ever be billed for. Astra's 54.7 is attached to a fixed capability; Mistral's 38.4 is attached to a moving one, and both of those facts cut in favour of the buyer in a different way.

Where Mistral's numbers come from, and who ran them

Mistral's launch material is unusually careful about provenance, and the carefulness is itself information. Three of the headline agentic-coding figures — 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0, combining into a 49.8% Coding Agent Index score — are described by Mistral as evaluated privately by Artificial Analysis ahead of that harness' public launch. They are not on the public index yet, which makes them stronger than a self-run benchmark and still not reproducible by you today.

The strongest independent datapoint in the release is one Mistral did not run. Surge AI, a third-party annotation firm, ran a blind human evaluation on coding quality with model identities hidden, and professional annotators rated Mistral Large 4 Preview at 3.74 on a 1–5 scale — second of five, behind only Claude Opus 5 at 4.22, and ahead of Kimi K3 at 3.59, GLM-5.3 at 3.60 and GLM-5.2 at 3.40. A blind human eval with a named lab on the other end is a different kind of evidence from a self-published table, and it is the single strongest independent result in the announcement.

Everything else in Mistral's table is vendor-reported and unreproduced, including the cyber figures below. Read them as claims a confident lab makes, not as measurements.

What Mistral is actually selling: the refusal

This is the part of the comparison no benchmark captures, and it is the sharpest thing either vendor has said in this matchup.

On Artificial Analysis' Cyber Index, Mistral says Mistral Large 4 ranks in the global top five and leads open-weight models developed outside China. On one specific test — reproduce a real vulnerability in open-source software and then patch it — it reports 82%, described as the highest of any model, along with 93% of Cybench's 40 competition exercises. Then it makes the observation that gives the release its edge: several leading closed models, Claude Opus 5.5 and GPT-6 Astra named among them, score near zero on that same test because they refuse the task.

Both halves of that are vendor-reported. Neither has been independently reproduced. But the mechanism is not exotic and it is not a swipe — OpenAI shipped GPT-6 Astra behind what coverage at launch described as a Critical cyber threshold, and a model with a tighter refusal posture will decline a vulnerability-reproduction request that a model with a looser one will attempt. The 82% and the near-zero are two sides of the same policy decision, and which one you want is a function of what you do for a living. A red team evaluating its own software wants the 82%. A regulated platform serving end users may specifically want the refusal, and may be willing to pay $10.00 and $50.00 per million tokens for it.

That is the difference between these two models that no leaderboard row will ever show you, and it is the reason they will coexist in the market rather than one displacing the other.

What "open-weight" means this week

Three caveats belong on Mistral Large 4's side of the sheet, and each one has a price attached.

• The weights are not out. There are no checkpoints, no licence file and no repository today. Mistral says they land by the end of October 2026 after a red-teaming window in which vetted partners and national authorities get an unlocked build. Until then the only way in is the preview API, which means the "run it on your own hardware" argument is a roadmap item, not a capability.

• The independence flag does not match the marketing. Artificial Analysis lists the model as isOpenWeights: false with the categorisation "proprietary," because a preview served from the vendor's own cluster is what exists. That flag is the cleanest single illustration of the gap between the words "open-weight" and the artifact.

• The price you are quoted is promotional. Mistral's card shows $0.68 and $2.09 against a struck-through list of $1.36 and $4.18. Model your bill on the latter. A preview is also a stage label in the contractual sense — the model can be re-pointed, re-priced or retired with less notice than a GA release, so a pipeline tuned on this week's build carries a re-validation cost you did not budget for.

None of that makes the preview a bad buy. It makes it a buy with an expiry date and a re-validation bill, which is a different kind of cost from a per-token rate.

Sovereignty is a procurement line, not a slogan

Mistral trained Mistral Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacentres, raised €3 billion in a Series D specifically to build that capacity, and is serving the preview from the same infrastructure. For a regulated buyer this is not a marketing position; it is a requirement GPT-6 Astra cannot currently satisfy at any price. OpenAI's model is reachable through OpenAI's own API, Azure and AWS Bedrock — all of which are excellent for scale and none of which is a European-operated, end-to-end pipeline.

The second half of the sovereignty argument lands when the weights do. An organisation with the GPUs can then run a 1.05-trillion-parameter sparse model on-premises with no provider in the loop, which changes both the data-residency story and the refusal behaviour at the same time. Astra will still be an API call when that happens. If your roadmap has an on-premise phase, Mistral Large 4 is the only one of these two that has one — and if it does not, that entire column is worth nothing to you.

Running both without betting on either

The awkward shape of this matchup for a router is that one side is a model we serve and the other is a preview we do not. GPT-6 Astra is in the OrcaRouter catalogue on OpenAI's list price with nothing added, on the same OpenAI-compatible key as 200-plus other models, with automatic failover across upstream providers — which matters here because a preview is exactly the kind of dependency you want off your critical path until it settles. Mistral Large 4 is not routed by us while it is served from Mistral's own cluster behind a preview flag; the honest version of our part of this comparison is the Astra side of it, plus the ability to run the same task against either without a second contract. When a generally available Mistral endpoint appears, it inherits the same 0% passthrough the rest of the catalogue runs on, and a vendor price change lands on the same key the day it happens rather than at the next renewal.

What one key buys you today is the A/B itself: the same prompt, the same evaluator, two model strings, one bill — which is how you answer the only question that matters about a preview, namely whether its refusal posture and its long-context recall are worth sixteen index points to your workload.

The decision rule, and the date that changes it

If the work is hard reasoning, agentic coding, or anything measured by the top of the Intelligence Index, GPT-6 Astra is ahead and the price gap is the premium for it — $10.00 and $50.00 per million tokens, repricing to $20.00 and $75.00 on requests above 272,000 input tokens.

If the work is European-resident by requirement, refusal-sensitive, or headed for on-premises hardware, Mistral Large 4 is the only one of the two that answers the question at all, at a quarter to a seventh of the token price, with the caveat that the artifact you deploy later will not be the artifact you evaluate now.

The date to watch is the end of October 2026, when Mistral says the checkpoints land. Three things change on that day: the licence gets a name, the self-hosting argument stops being a promise, and the launch price attached to a moving preview stops being the price attached to a fixed one. Anyone currently comparing these two models on capability should re-run the comparison after that, because one column of it will have moved.