Mistral Large 4 added to BenchGecko · specs, launch pricing, vendor benches

What landed

Mistral Large 4 by Mistral AI entered public preview on 2026-10-06. BenchGecko lists the row as mistral-large-4-0 with list pricing from Mistral's API docs.

The model card describes a Mixture-of-Experts multimodal model with 52B active parameters of 1.05T total, plus a 1.6B vision encoder. Context is 1M tokens (1,048,576 on BenchGecko).

Pricing

Mistral's pricing page shows a launch sale on Large 4:

• Input · $0.68 per M tokens (list $1.36)
• Cached input · $0.07 per M tokens (list $0.14)
• Output · $2.09 per M tokens (list $4.18)

BenchGecko currently stores $0.68 input and $2.09 output per M tokens with a 1,048,576 context window for Mistral Large 4.

Vendor-reported benchmarks

Figures below are from Mistral's 2026-10-06 announcement. They are vendor-reported, not BenchGecko's own tests.

Coding and agents

• DeepSWE v1.1 · 61.7%
• SWE-Atlas QnA · 59.4%
• Terminal-Bench 4 · 28.3%
• Combined Coding Agent Index · 49.8%
• AutomationBench · 59.9%
• AA-Briefcase · 1,393 Elo

A Surge AI blind human coding eval (1 to 5 scale) put ML4 Preview at 3.74, behind 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.

Cyber and safety

• Cybench · 93%
• AA Cyber Index vulnerability reproduce-and-patch task · 82%
• Lakera B3 attack resistance · 93.3%

Multimodal

On Dense 200 visual grounding, Mistral reports 42% for ML4 versus 41% for GPT-6 Astra.

Training notes

Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters, with multilingual training data spanning more than 160 languages. Weights are planned for end of month. The announcement frames ML4 as the first milestone after a €3 billion Series D.

Sources

• Introducing Mistral Large 4
• Mistral Large 4 model card
• Mistral API pricing
• BenchGecko model page: /model/mistral-large-4-0

Source: BenchGecko