Mistral AI's Mistral Large 4 is the company's first trillion-parameter model - 1.05T total parameters with 52B active in a mixture-of-experts layout - opened to public preview on October 6, 2026. Trained in European datacenters, it carries a 1M-token context window, native image understanding and support for 160-plus languages. List price is $1.36 / $4.18 per million tokens, dropping to $0.68 / $2.09 during preview; open weights are promised by the end of October. Mistral positions it as the open-weight counterpunch to gated US frontier models. Self-reported benchmark figures (Cybench 93%, CyberGym-E2E 82%, DeepSWE 61.7%) are vendor numbers, not yet independently verified.
Parameters
1.05T total, 52B active (MoE)
Context Window
1M tokens
License
Mistral License (weights promised end of Oct 2026)
Release Date
2026-10-06
API Pricing
Input Price (per 1M tokens)
$0.68
Output Price (per 1M tokens)
$2.09
Billing Mode: preview ($1.36/$4.18 list)
Strengths
- •First European trillion-parameter model, trained in EU datacenters - a sovereignty story competitors cannot copy
- •1M-token context with native image understanding across 160-plus languages
- •Open-weight line: weights promised by end of October, directly countering gated US frontiers
- •Aggressive preview pricing at $0.68 / $2.09 per million tokens
Weaknesses
- •Launch is preview-only; open weights not yet released, so 'open' is a promise not a fact
- •Benchmark figures (Cybench 93%, etc.) are self-reported; partly a policy story since US rivals refuse cyber tasks
- •No independent third-party verification of the 1.05T / 52B claims yet
- •List price $1.36 / $4.18 is higher than many mid-tier rivals once preview ends
Use Cases
- •European-regulated enterprises needing EU-trained, sovereignty-aligned large models
- •Long-document and multimodal agentic workflows within a 1M-token window
- •Multilingual content and image-understanding tasks across 160-plus languages
Deep Analysis
Parameters
1.05T total / 52B active (MoE)
trained in EU datacenters
Context window
1M tokens
multimodal (text + image)
Introductory pricing
$0.68 / $2.09 per 1M
list $1.36 / $4.18; weights Oct 2026
CyberGym-E2E
82%
autonomous vuln reproduce/patch
Surge AI human rating
3.74 / 5
vs Opus 5 4.22
Strengths
- ・First European trillion-parameter model, trained from scratch in EU datacenters on 3,800 GB200 GPUs
- ・Open weights promised by end of October 2026, giving EU teams a home-grown large model they can self-host
- ・1M-token context and native image understanding across 160-plus languages
- ・Strong vendor-reported cyber scores (Cybench 93%, CyberGym-E2E 82%) and high refusal rate on harmful cyber requests
Weaknesses
- ・Intelligence Index 38.4 (AA) sits below the frontier tier despite the 1T scale
- ・Key cyber scores are partly a policy story - US rivals refuse those tasks outright
- ・License for the upcoming weights is not yet published; self-hosting is not yet possible
- ・Vendor-reported figures, unverified at launch
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| GPT-6 Astra | frontier | n/a | 96 | n/a |
| Claude Opus 5.5 | frontier | n/a | n/a | $4 / $20 |
| GLM-5.3 | open | 61 | n/a | ~$1.36/$4.18 |
Mistral AI's flagship frontier model, opened in public preview on October 6, 2026. A 1.05T-parameter mixture-of-experts (52B active) trained from scratch in European datacenters, with a 1M-token context, native image understanding and 160-plus languages. Weights are promised by the end of October.
Sources
Analysis generated: 2026-10-08