Mistral Large 4 Enters the AI Arena: October 2026 Rankings

👁 117 reads

Two new model variants joined the AI writing Arena this week: Mistral Large 4 and Mistral Large 4 (non-reasoning). Mistral Large 4 entered highest, taking 17th place among 26 cloud models with an average score of 6.95 out of 10. Claude Opus 5.5 remains unchanged at the top with 8.94/10.

The weekly Arena report tracks changes in the AI model Arena: who arrived, who shifted positions, and why. I calculate the ranking from the average scores given by LLM judges across 9 writing challenges (conducted in Polish), displayed alongside the Lab benchmark score (0-100), which also evaluates text proofreading and linguistic nuance tests.

New entries to the Arena

Mistral Large 4: 17th place

• Judges’ average: 6.95 out of 10 across 8 tasks.
• Lab benchmark score: 69.2 points out of 100.
• Best performance: newsletter translation - 8.3/10, ranking 10th of 26.
• Weakest performance: blog article - 5.5/10, ranking 19th of 26.
• Unrated (empty response): product description - this task is excluded from the Arena average.

Read the complete evaluation with error breakdowns: Mistral Large 4 review and test results.

Mistral Large 4 (non-reasoning): 19th place

• Judges’ average: 6.50 out of 10.
• Lab benchmark score: 65.6 points out of 100.
• Best performance: customer complaint reply - 7.5/10, ranking 14th of 26.
• Weakest performance: explaining inflation to a child - 5.5/10, ranking 25th of 26.

Read the complete evaluation with error breakdowns: Mistral Large 4 review and test results.

What the Mistral results tell us

Mistral Large 4 is the first new release added to the Arena since its launch on October 2. Checking the Lab benchmark records confirms no other models debuted during this window; the preceding runs (Llama 4 Maverick, Llama 4 Scout, and Muse Glimmer 30B on September 30) were included from day one.

• On writing quality, Large 4 clearly outperforms its predecessor. Mistral Medium 3.5 sits at 25th place with an Arena average of 5.13, while reasoning-enabled Large 4 hits 6.95 at 17th place. The biggest improvements show up in text summarization (7.8 vs 3.9) and customer complaint handling (7.0 vs 4.1).
• In the Lab benchmark, the gap almost vanishes: 69.2 points versus 68.0. The Lab score factors in strict rule compliance and complex proofreading. On the challenging proofreading task and product description, Large 4 with reasoning exhausted its output token limit and returned blank outputs. The Arena ignores empty outputs, but the Lab scores them as outright failures.
• The non-reasoning build is cheaper and more consistent. It scored an average of 6.50 for 19th place, successfully completed all 9 writing tasks, and ran all 13 Lab benchmark tasks for $0.015 instead of $0.134.
• It remains far behind the top tier. Claude Sonnet 5.5 scores 8.28, GPT-6.1 Sol achieves 8.57, and even the budget-friendly Gemini 3.8 Flash hits 8.09. However, Large 4 beats GLM-5.3, Claude Haiku 4.5, and both Llama 4 variants.

For specs, pricing details, and open weights, see Mistral Large 4: release, specs, pricing, and weights. For cost comparisons against rival models, check Mistral Large 4 vs Claude, GPT, Gemini, and DeepSeek. You can also explore the company overview in our Mistral AI guide.

Risers and fallers

Every downward move this week was purely mechanical: 8 models slipped down the leaderboard simply because the two new Mistral variants scored higher. Their underlying test scores did not drop.

Minor score adjustments occurred without position changes: Claude Opus 5.5 (-0.04) and GPT-6 Luna (-0.03). This stems from rescording the baseline anchor models during the new run, shifting their cumulative average across all rounds by hundredths of a point.

Leaderboard standings

How the ranking is calculated

• Every model receives identical writing prompts: an article, product description, customer complaint response, LinkedIn post, short story, 10 SEO headlines, summary, explaining inflation to a child, and translation. Note that we tested these prompts in Polish.
• Outputs are evaluated blindly by LLM judges on three criteria: accuracy, natural phrasing, and prompt adherence. The average forms the 0-10 score.
• This leaderboard tracks cloud models. Local models (tested on an RTX 4090) are tracked separately due to calibrated scoring standards.
• Comparisons are made against the prior update. A lower position with an identical score simply indicates a newer model debuted higher.
• The Lab score (0-100) is calculated separately: 70% judge ratings, 20% hard prompt constraint adherence, 10% text cleanliness, plus specialized proofreading and linguistic tests.

What is next

New model rollouts from the AI launch calendar are added to the Lab benchmark the day API access goes live, then integrated into the Arena. You can inspect side-by-side model outputs in the Arena and compare image generation tools in the AI image generator arena.

// read moreMistral Large 4 vs Claude, GPT, Gemini and DeepSeek: PricingBenchmarksMistral Large 4 (Le Chonk): Specs, Pricing, and Open WeightsNewsMicrosoft-Decision-1: Ranking, Pricing and Test ResultsBenchmarks

Kinetyra