Mistral Large 4 vs Qwen3.8 Max: A Claimed Lead That the Published Numbers Deny

Mistral's launch post for Mistral Large 4 states that its combined Coding Agent Index score of 49.8% puts it "ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max". Artificial Analysis already has Qwen3.8 Max on the same index family, and its Terminal-Bench 4.0 figure is 38.9%. Mistral reports its own model at 28.3% on that harness. Both numbers come from Artificial Analysis — one public, one produced privately ahead of the harness launch — and they point in opposite directions. That is not a reason to dismiss either model. It is the most useful thing about this matchup, because it tells you exactly which of the two you can verify today, and it sets the price question up sharply: Qwen3.8 Max costs $2.00 per million input tokens and $6.00 per million output against Mistral Large 4's $0.68 and $2.09, roughly three times as much, and on every published measure it is the stronger model.

Both are Alibaba-scale and Mistral-scale attempts at the same thing — a natively multimodal flagship serving about a million tokens of context with reasoning, tools and structured outputs — released nine weeks apart, in public preview and generally available respectively. One of them has been shipping since 3 August 2026 and has a full column of independent scores. The other arrived on 6 October 2026 with a promised weight release by the end of the month and a reinforcement-learning run its vendor says is still climbing.

The published column and the promised one

Qwen3.8 Max is Alibaba's highest-capability tier to date, and Alibaba positions it directly against GPT-5.5, Claude Opus 4.7 and Gemini 3.1 Pro in its own migration guide. It is natively multimodal across text, images and video with text output, carries a 1,000,000-token context window, and its capability matrix confirms thinking mode, function calling, built-in tools and structured outputs. It is served in three wire formats — OpenAI-compatible, Anthropic-compatible and native DashScope — from Beijing, Singapore, Tokyo, Frankfurt and further regions.

On Artificial Analysis it scores 45.4 on the Intelligence Index v4.3.2, 17th of 147 models, and 76.2 on the AA Coding index, which is 9th of 138 and tied with Kimi K3 at the top of the open-weights field. It holds 92.8% on GPQA Diamond, 80.3% on long-context recall, 52.1% on SciCode and 47.8% on tau-banking. Its Terminal-Bench 2.1 result is 88.8%.

Mistral Large 4 is a granular mixture-of-experts model with 49 billion active parameters out of 1.05 trillion total and a 1.6-billion-parameter vision encoder, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres on data covering more than 160 languages. It takes text and images across roughly a million tokens of context. Its Artificial Analysis page exists under the title "Mistral Large 4 Preview" and carries no Intelligence Index score; the weights are promised by the end of October 2026.

• Intelligence Index — Qwen3.8 Max 45.4 (17th of 147) vs Mistral Large 4 not yet published

• Coding index — Qwen3.8 Max 76.2 (9th of 138) vs Mistral Large 4 none comparable

• Terminal-Bench 4.0 — Qwen3.8 Max 38.9% vs Mistral Large 4 28.3%, evaluated privately

• Context window — Qwen3.8 Max 1,000,000 tokens vs Mistral Large 4 about 1,000,000

• Input modality — Qwen3.8 Max text, image and video vs Mistral Large 4 text and image

• Input price — Mistral Large 4 $0.68 per 1M vs Qwen3.8 Max $2.00 per 1M

• Output price — Mistral Large 4 $2.09 per 1M vs Qwen3.8 Max $6.00 per 1M

• Cached input — Mistral Large 4 $0.07 per 1M vs Qwen3.8 Max $0.25 per 1M

• Weights — Mistral Large 4 open licence promised end of October 2026 vs Qwen3.8 Max closed

• Released — Qwen3.8 Max 3 August 2026 vs Mistral Large 4 6 October 2026, preview

Resolving the contradiction, as far as it can be resolved

The two Terminal-Bench 4.0 figures are genuinely inconsistent, and the most likely explanation is the least dramatic one: they are different runs. Mistral's 28.3% is described as evaluated privately by Artificial Analysis ahead of the harness's public release, which means it may have been measured on a pre-release harness, at a different reasoning effort, or with different sampling. Qwen's 38.9% is on the published harness at whatever configuration Artificial Analysis standardised. Comparing them directly is comparing a draft to a final, and the draft is the one that looks worse.

It is also worth noticing which way the contradiction cuts. Mistral claims a lead on the combined Coding Agent Index — 49.8%, ahead of Qwen3.8 Max — while publishing a Terminal-Bench 4.0 figure that trails Qwen's by ten points. A combined index aggregates several evaluations, so both can be true simultaneously, and a model with a strong DeepSWE v1.1 result (Mistral reports 61.7%) and a weak Terminal-Bench result can still win the average. But until the Coding Agent Index is public, the only component of that claim a reader can check is the one where Mistral is behind. That asymmetry is the correct thing to be suspicious of, in either direction.

What three times the price actually buys

Two things, and only one of them is performance.

Video is the concrete one. Qwen3.8 Max accepts video natively as a first-class input; Mistral Large 4 does not. Mistral's multimodal chapter is about images — grounding dense scenes, verifying mechanical parts in technical drawings, retrieving evidence from PDFs, scanning gigapixel satellite imagery — and none of that requires moving pictures. If your pipeline processes recordings, demonstrations, or video archives, the price comparison is irrelevant, because one of these models cannot take your input. This is a capability boundary, not a rate-card difference, and no reprice will move it.

The second is proof. Qwen3.8 Max's 45.4 Index score, its 9th-place coding rank and its 88.8% on Terminal-Bench 2.1 are all reproducible on published harnesses by anyone who wants to spend the tokens. Mistral's ML4 offers a higher claimed ceiling on agentic coding and automation, a blind human evaluation in which it edged Kimi K3 and GLM-5.3 on perceived coding quality, and a specific cybersecurity claim — 82% on Artificial Analysis's reproduce-and-patch test, the highest of any model tested, with 93% on Cybench. Every one of those is vendor-reported or privately evaluated. A team that has to justify a model choice to a procurement review pays the three-times premium to have the answer in writing.

That said, the cybersecurity figure is the one claim in this comparison that a lower price cannot offset, because it is a capability, not an advantage. Mistral reports that several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same reproduce-and-patch test because they refuse the task, and that ML4's average refusal rate on cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm is higher than that of other open-weight models while still completing the work. Qwen3.8 Max is not reported on that test. If your work involves vulnerability reproduction, the question is not who is cheaper; it is who will do the job at all.

Weights, and where they run

This is where Mistral Large 4 holds the one structural advantage that Qwen's three-times premium cannot touch. Mistral has committed to releasing ML4's weights under an open licence by the end of October 2026, alongside details of the architecture and post-training methodology. That converts a per-token cost into a capital decision: if you have the GPUs, you stop paying per token, stop sending data to a third party, and stop being subject to someone else's moderation policy. Alibaba's Max tier is a closed, API-only offering, and nothing suggests that is changing.

Geography is the same argument in a different register. Qwen3.8 Max is served from Beijing, Singapore, Tokyo and Frankfurt among others — a genuinely global footprint, but one operated by Alibaba. Mistral Large 4's public preview runs on Mistral's own European infrastructure, in a European region Mistral operates end to end and independently of other digital service providers, under European law. For a European regulated buyer, those two sentences support very different procurement outcomes even though both models are available in the same city. Mistral's €3 billion Series D — the largest equity round ever raised by a European technology company — exists to fund exactly that distinction, and ML4 is its first deliverable.

Routing across two shapes of availability

The unusual thing about this pairing is that the two models differ not just in price and capability but in how you reach them. Qwen3.8 Max speaks OpenAI-compatible, Anthropic-compatible and native DashScope protocols from a spread of regions; Mistral Large 4 is a single preview endpoint from one operator. Normalising that from your application code is real work, and it is the kind of work that starts out small and grows every time a provider changes something.

Both sit behind one OpenAI-compatible endpoint on OrcaRouter with 0% markup, which is what makes the price gap in this article stable rather than a snapshot: provider list prices pass through unmodified, so the day Mistral ends its preview rate or Alibaba moves Qwen's, the figures your routing rules evaluate against move with them. The routing DSL is where the two models' different strengths become one call surface — video-bearing requests to Qwen3.8 Max, image and document work to Mistral Large 4, and anything that has to hold a strict budget sent to ML4 first with Qwen as the escalation when the combined-index claim turns out to matter on a real task. Automatic failover is worth more in this pairing than most, because one leg is a preview API in its first month and the other is a closed tier you cannot capacity-plan. Where the 49.8-versus-76.2 coding question is genuinely undecided for your workload, model fusion can run both on the same prompt and let a panel choose, which is cheaper than running a bake-off twice.

Verdict

Qwen3.8 Max is the better model by everything that has been published. Its 45.4 Index score, 76.2 coding index and 38.9% on Terminal-Bench 4.0 are on the same revisions as everybody else's, it takes video, it has been shipping for two months, and it is reachable through three wire formats from five regions. If performance and verifiability are the deciding factors, the three-times price is what a shipped, measured flagship costs.

Mistral Large 4 wins on price by roughly three times, on cache reads by more than three times, on cybersecurity work where the alternative refuses, and on the two arguments Qwen cannot answer: promised open weights by the end of October, and a European deployment Mistral runs end to end. The 49.8% Coding Agent Index claim is the thing that would settle it — if that number lands on a public harness ahead of Qwen's, the cheaper model becomes the better one and the premium disappears. If it does not, Mistral is selling a preview price on a model that trails a released competitor, and the honest position is that you are buying the roadmap rather than the model.

Watch three dates. The end of October 2026, when ML4's weights either ship or the open-weights claim softens. The publication of the Artificial Analysis Coding Agent Index, where Mistral's private evaluations and Qwen's public scores meet on one column. And the end of the preview period, whenever Mistral announces it, because $0.68 and $2.09 is a launch rate and no forecast should assume it survives the weight release.