Mistral Large 4 vs Claude Opus 5.5: The Only Head-to-Head Number Is Against the Wrong Version

The one blind head-to-head that exists for Mistral Large 4 and Claude Opus 5.5 was measured against the previous version of one of them. Surge AI hid the model identities and had professional annotators rate coding output on a 1–5 scale; Mistral Large 4 Preview finished second of five at 3.74, behind Claude Opus 5 at 4.22 and ahead of Kimi K3 at 3.59 and GLM-5.3 at 3.60. Claude Opus 5.5 shipped on 22 September, so the closest thing to a published human comparison for this pairing is three weeks older than the opponent's current release, and nobody has re-run it. Everything else is comparable, and it points in one direction: on Artificial Analysis's Intelligence Index, Mistral Large 4 Preview reads 38.4 and Claude Opus 5.5 reads 58 — first of 216 models — while the cost of completing a single index task is $1.13 against $5.98.

Those two facts are the whole shape of the decision. The capability gap is wide and, as of this version, wider than it was. The cost gap is a little over five times, and it did not move when Anthropic cut its prices.

What Claude Opus 5.5 is, on Anthropic's own account

Claude Opus 5.5 is Anthropic's current flagship, released 22 September 2026 and positioned at long-horizon agentic work: multi-step sessions, heavy tool use, and code. It lists at $4 per million input tokens and $20 per million output, with cache reads at $0.20 and cache writes at $5 — Anthropic describes those as 20% below Opus 5 on input and output and 60% below it on cache reads. There is a Fast mode at $8 / $40 that the company says runs up to 2.5x quicker inside Claude Code and the Claude Platform. It is available on Amazon Web Services, Google Cloud and Microsoft Azure, with zero data retention offered, and Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.

Two details on that page matter more than the rest. The first is a migration constraint rather than a feature: Anthropic states that Claude Opus 5.5 is no longer available with thinking mode switched off. If your current pipeline calls Opus 5 in non-thinking mode for cheap, fast, deterministic turns, that configuration has no direct successor in this model. The second is Anthropic's own caveat, printed beside its scores, that benchmark margins have become a less reliable guide to real-world differences — which is an unusual thing for a vendor to write next to a table it is winning, and it is worth taking at face value when you read the numbers below.

The vendor-reported figures from Anthropic's announcement are Terminal-Bench 4.0 at 66.4%, FrontierCode v1.1 at 54.4%, CursorBench 4.0 at 57.8%, GDPval-AA v2.1 at 1,846 Elo, AutomationBench at 40.0%, Humanity's Last Exam at 67.7% with tools, Terminal-Bench-Science 0.1 at 58.7% and OSWorld 2.1 at 81.8% partial. All of those are Anthropic's own runs, with the comparison models' scores taken from the vendors' own pages, and none of them has been reproduced independently for this pairing or any other.

What Mistral Large 4 is, as of 13 October

Mistral Large 4 is a public preview announced on 6 October 2026, revision v26.10, and it is the largest model Mistral AI has shipped. The company describes it as a natively multimodal sparse mixture-of-experts with 1.05 trillion total parameters and 52 billion active, plus a 1.6-billion-parameter vision encoder, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, on data spanning more than 160 languages. Artificial Analysis records the active count as 49 billion rather than 52 billion — a three-billion gap on one dimension, noted rather than averaged. Mistral's own model card states a 1M-token context window; Artificial Analysis reads 524,288 on the configuration it tests, and its FAQ says 520k, which is the difference between a ceiling a vendor advertises and a window a served deployment actually hands you.

Pricing is $1.36 per million input tokens, $0.14 cached input and $4.18 per million output on the rate card, with a launch promotion showing the same three lines struck through at $0.68, $0.07 and $2.09. The status labels on the card read Public Preview, Open and v26.10 — and the second of those is the one to read carefully. Mistral says the weights will be released by the end of the month, and the launch post repeats that commitment twice; what it does not do is name a licence. Artificial Analysis tracks the field independently and records the licence name as null, marks the model isOpenWeights: false, and classifies it as proprietary. Nobody is contradicting anybody there. A preview served from the vendor's own cluster, with no checkpoint repository and no licence text, is a callable model rather than an open-weights release, and the open-weight claim is a plan with a deadline attached to it rather than a download link.

Mistral's own benchmark figures, which are vendor-reported throughout, are 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4.0 and a combined Coding Agent Index of 49.8%; 59.9% on AutomationBench across 657 business workflows; 1,393 Elo on AA-Briefcase; 42% on Dense 200 for visual grounding; and, on its cyber evaluations, 82% on a reproduce-and-patch test Mistral describes as the highest of any model, alongside 93% on Cybench's 40 competition exercises. The company also commissioned a blind human study and published it, which is a stronger form of evidence than a self-run benchmark and still a study whose publication it controlled.

What the 5.5 bump changed about this decision

If you read the Mistral Large 4 versus Claude Opus 5 page on this blog, three things here are new, and only one of them is the price cut.

• The capability gap got wider, not narrower. Claude Opus 5 scored 50.8 on the Intelligence Index; Claude Opus 5.5 scores 58. Against Mistral Large 4 Preview's 38.4, the distance goes from 12.4 points to 19.6. A version bump that adds seven points to the top of the board while the challenger stands still makes the frontier lane a clearer buy, not a murkier one.

• The price cut did not reach the bill. Anthropic cut input and output 20% — $5 / $25 to $4 / $20 — but Artificial Analysis's cost to complete one index task went up, from $5.86 for Claude Opus 5 to $5.98 for Claude Opus 5.5. The cheaper tokens were spent by a model that used more of them. That is the single most useful number on this page and it is worth sitting with: list prices are an input to your invoice, not a prediction of it.

• One configuration disappeared. Thinking mode can no longer be switched off in Opus 5.5. If you were buying Opus 5 partly for a cheap non-thinking path, this model does not replace that path, and the replacement decision is now a two-model one rather than a version swap.

What did not change: the cache story still favours Anthropic for repeated context — $0.20 per million on a cache read against Mistral's $0.14 list and $0.07 promotional. On a long-context agent loop with a stable prefix, that is the row that moves the bill most, and both numbers need the same scrutiny the headline prices just got.

The two most verbose models on the board

Artificial Analysis publishes an output-token count for each model's index run, and the two models here are the extreme ends of it. Claude Opus 5.5 generated 260 million output tokens on that run. Mistral Large 4 Preview generated 200 million. The median across the models measured is 85 million. Both of them spend two to three times what a typical model spends to answer the same questions, and only one of them charges $4 and $20 per million while doing it.

Verbosity is not a defect, and it is not a quality signal either. It is a multiplier on a price, and it is the reason the Opus price cut landed without reaching the per-task cost. If you are budgeting from the rate card alone, budget the multiplier too: at $20 per million output, a model that emits three times the median has an effective output rate closer to $60 per million against a median-behaving model, and 260 million tokens over an evaluation suite is not a rounding error.

The scoreboard

• Price per million tokens, input / output — Mistral Large 4 $1.36 / $4.18 list, $0.68 / $2.09 promotional, vs Claude Opus 5.5 $4.00 / $20.00

• Cached input — $0.14 list, $0.07 promotional vs $0.20 per million read and $5.00 per million write

• Cost per Intelligence Index task — $1.13 vs $5.98, per Artificial Analysis

• Intelligence Index — 38.4, 65th of 216 vs 58, 1st of 216, per Artificial Analysis

• Output tokens on that index run — 200M vs 260M, against a median of 85M

• Context window — 1M stated by Mistral with 524,288 on the tested configuration vs 1M on Artificial Analysis's reading of Claude Opus 5.5, whose announcement page states no figure either way

• Maximum output — not published by either vendor

• Measured speed — not published for Mistral Large 4 Preview vs 95.2 output tokens per second, with a $8 / $40 Fast mode Anthropic says reaches up to 2.5x

• Weights and licence — promised by end of October with the licence unnamed, currently flagged proprietary by Artificial Analysis, vs closed

• Availability — Mistral's preview API and Mistral Studio, served on Mistral's own European infrastructure, vs AWS, Google Cloud, Azure and the OrcaRouter catalogue

When Mistral Large 4 is the right pick

Pick Mistral Large 4 when the work is high-volume, well-specified and cheap to check. Classification, extraction, summarisation at scale, structured-output generation, first-pass code review, and the long tail of internal tasks where a wrong answer costs a retry rather than an incident. At $1.13 per index task against $5.98, a pipeline that runs a million tasks a month does not need a subtle argument to justify the cheap model; it needs the cheap model to be good enough, and on AutomationBench — the row where a model is asked to complete ordinary business workflows — Mistral's own reported 59.9% is competitive with the figure Anthropic reports for Opus 5.5 on the same harness. Different runs, different vendors, so read it as a signal rather than a result, but it is the row closest to what a business unit actually asks a model to do.

Mistral Large 4 is also the one side of this pair with a credible route to self-deployment, and that is worth more than a benchmark point to some readers. A promise of weights by the end of October is not weights, and an unnamed licence is not a licence, so this is a reason to watch rather than a reason to migrate today. But among the two models here it is the only one where a data-residency or mid-incident-access argument can end in "we run it ourselves". The European training and serving story sits on the same side of that ledger.

When Claude Opus 5.5 is the right pick

Pick Claude Opus 5.5 when the task is long, the steps depend on each other, and being wrong is expensive. Terminal-Bench 4.0 is the sharpest dividing line in this comparison — Anthropic reports 66.4% for Opus 5.5 against Mistral's 28.3% for its own model on the same harness — and terminal work is the closest public proxy for agentic coding, which is most of what teams are buying a frontier model for. A 19.6-point Intelligence Index gap and a near-2.4x Terminal-Bench gap put the same way: this is not a close call on the hard end, and no price argument closes it if your workload lives there. Add the 1M context, the measured 95.2 tokens per second against an unpublished figure on the other side, the Fast mode for latency-sensitive paths, and availability across three major clouds with zero data retention, and the premium buys real operational slack.

The practical question is whether you need the whole premium on every call. Most teams running a frontier model at volume find that a minority of requests — the ones with genuine multi-step ambiguity — consume most of the capability, and that the rest do not. That is a routing decision, not a model decision, and it is the case where running both is the actual answer. Claude Opus 5.5 is in the OrcaRouter catalogue at Anthropic's list price with 0% markup on the same OpenAI-compatible endpoint as 200-plus other models, so trying both costs one key rather than a second vendor contract. Mistral Large 4 is not in our catalogue today — it is reached through Mistral's own preview API and via several third-party platforms — and that is stated as a fact about our list, not a soft launch announcement.

For a preview, the feature that earns its place is failover. Sending a slice of traffic to a model whose licence, weights and served context are all still open questions means the useful failure mode is a reroute rather than an outage, and you learn the model's real failure modes on your own data instead of on somebody's leaderboard.

The one fact that decides it

Nobody has benchmarked Mistral Large 4 and Claude Opus 5.5 head to head at matched settings on a named harness — the only blind human number that exists is against Claude Opus 5, and the version you are choosing against is not Claude Opus 5 — so the honest basis for a decision today is that the cheap model costs a fifth as much per completed task and the expensive one is roughly twenty index points and a near-2.4x Terminal-Bench gap ahead of it, at a list price Anthropic cut without cutting the per-task cost.

If a failed answer in your workload costs less than 4.6 retries, buy Mistral Large 4. If it costs more, buy Claude Opus 5.5 — and if you cannot answer that question about your own traffic yet, run both on one endpoint and let a month of real requests answer it for you.