Reviewed tool profile

Mistral Large 4

Mistral’s 1T-parameter multimodal MoE model for coding, agents, cybersecurity and document work.

Sources checked Oct 7

Pricing: API usage is billed per million tokens in USD. The listed price is $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. The model page currently shows a 50% sale price: $0.68 input, $0.07 cached input and $2.09 output. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
Available as: Available through the Mistral API and Mistral Studio, the developer console with keys, a Playground, evaluations, agents and SDKs. The model supports the chat completions, conversations, agents and batch endpoints. · Mistral says the model will be offered in multiple regions, including a European deployment it operates end to end. It says organisations will be able to run it on private cloud or on-premises once weights are available. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗ Official source 1 ↗
Primary group: Coding & Development

What it does

Mistral Large 4 (nicknamed “le Chonk”) is Mistral AI’s largest model, released as a public preview on October 6, 2026. It is a multimodal Mixture-of-Experts model with 1.05 trillion total and 52 billion active parameters, a 1.6B vision encoder and a 1M-token context window. It combines instruction, reasoning and agentic use in one model and is available now through the Mistral API in Mistral Studio; Mistral says open weights follow by the end of October 2026.

Official source 1 ↗ · Official source 2 ↗

Features

Agentic coding. Works across software engineering, repository understanding and terminal workflows. Mistral reports 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. Official source 1 ↗
Agents and tool use. Runs general-purpose agents that gather information, call tools and produce deliverables. The API supports function calling, structured outputs, built-in tools and the Agents endpoints. Official source 1 ↗ · Official source 2 ↗
Multimodal documents and images. Accepts image input and reasons over complex documents, charts and natural images, including visual grounding in technical drawings, PDFs and large geospatial images. Official source 1 ↗ · Official source 2 ↗
Cybersecurity work. Mistral positions it for defensive security tasks; in internal testing it analysed malware, prioritised vulnerabilities and wrote detection rules. Official source 1 ↗
Knowledge-work deliverables. Creates, edits and fixes spreadsheets and documents, and produces slides and PDFs for finance and legal workflows. Official source 1 ↗
Multilingual. Mistral says it is natively fluent in more than 160 languages. Official source 1 ↗

Pricing & free limits

Mistral API — list price

$1.36 input / $0.14 cached input / $4.18 outputUSD per million tokens

Pay-as-you-go usage through Mistral Studio. Input and output tokens are counted separately. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗

Mistral API — current sale price

$0.68 input / $0.07 cached input / $2.09 outputUSD per million tokens

Sale price shown on the Mistral Large 4 model page at the time of checking; the end date is not stated. Official source 1 ↗

Billing details

Mistral’s pricing page says batch processing halves the price and cached input tokens cut input cost by up to 90% for repeated prompts. Official source 1 ↗

Free access & limits

Free access and its limits were not established by the checked official sources.

Best for & limitations

Run coding agents on repository-level software engineering and terminal tasks. Official source 1 ↗

Support security teams with malware analysis, vulnerability prioritisation and detection-rule writing. Official source 1 ↗

Extract and verify information from PDFs, charts, technical drawings and satellite imagery. Official source 1 ↗

Build or fix spreadsheets and documents for finance and legal work. Official source 1 ↗

Write code for scientific data analysis, modelling and simulation. Official source 1 ↗

Public preview, still changing. The model is a public preview and Mistral says its reinforcement-learning run is still in progress, so behaviour and scores may change. Official source 1 ↗
Weights not yet released. At launch only the hosted API is available; Mistral says weights will be released by the end of the month. Official source 1 ↗
Higher cyber refusal rate. Despite its security focus, Mistral reports a higher refusal rate on malicious cyber prompts than other open models. Reduced-moderation access is limited to vetted partners during red-teaming. Official source 1 ↗