Mistral has opened public API preview for Mistral Large 4, its newest flagship model, allowing enterprise teams to test multi-step and cross-application tasks before a fuller release later in October. The company says downloadable weights and additional technical details will arrive by the end of the month. The preview began Oct. 6.

The release arrives as enterprises move beyond simple chatbot prompts toward agentic workflows that span email, customer relationship management, coding environments, and security operations. Mistral Large 4, nicknamed Le Chonk, is being positioned for those longer-horizon jobs. Its public scores cover business software, software development, and cybersecurity. But the company's benchmarks are only the starting point. Buyers must weigh performance against control, deployment options, and cost.

What enterprise testers can evaluate now

Mistral Large 4 is available through the company's API for tasks that require several steps instead of a single prompt and response. According to Mistral, early results include:

• Business software: 59.9% on AutomationBench, which covers 657 workflows using applications such as Gmail and Salesforce. The benchmark measures whether an AI system can complete a sequence of tasks across workplace software.
• Coding: 49.8% on the Coding Agent Index and 61.7% on DeepSWE v1.1. Both evaluations test how well an AI system works through multi-step software-development jobs.
• Cybersecurity: 82% on an evaluation that involves reproducing a real software vulnerability and then repairing it. Mistral also reported 93.3% resistance on a test designed to trick an AI system into following malicious instructions.

Professional work is part of the pitch as well. The model can edit spreadsheets and documents and work with visual material. Its AA-Briefcase result reached 1,393 Elo on an evaluation of extended knowledge-work assignments.

Those numbers suggest a model that can handle more than conversation. They do not, however, guarantee production readiness. For security work in particular, a high score on vulnerability reproduction and repair does not automatically mean that an AI-generated patch will preserve normal application behavior. Research into AI-generated vulnerability fixes has found that closing an exploit may not prevent regressions or unexpected side effects. Companies testing Large 4 for security tasks should validate generated fixes against expected application behavior before allowing automated changes in production.

Architecture, training, and self-hosting plans

Mistral trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in data centers across Europe. Preview API traffic runs on infrastructure operated by Mistral, which also plans a regional deployment governed by European law. Regional infrastructure is part of the company's broader AI strategy, giving European enterprises a hosted option that aligns with local regulatory expectations.

More consequential for some buyers is the planned release of downloadable weights. If Mistral follows through, security organizations and other enterprises could run the model in a private cloud or on their own premises under internal policies. Companies handling sensitive code or proprietary documents could keep greater control over where AI work runs. That control comes with trade-offs. Operating a model internally brings hardware, staffing, and maintenance costs that a hosted API keeps with the provider. Some information needed to plan an internal deployment, such as architecture details and post-training information, is still pending. Teams tracking the open-weight AI market can test the hosted version now, but sizing their own infrastructure will depend on the remaining technical details.

Benchmark claims and the limits of early scores

Mistral's benchmark results are important because they frame what the model is designed to do. AutomationBench, for example, is intended to measure multi-step workflows across business applications. A score of 59.9% means the model completed some but not all of those workflows. In coding, the Coding Agent Index and DeepSWE v1.1 test multi-step development tasks; scores of 49.8% and 61.7% indicate meaningful capability but also room for error. The 82% cybersecurity result and 93.3% prompt-injection resistance are notable, but they are measured on specific evaluations, not on every possible production scenario.

Enterprise teams should treat these results as a starting point for their own testing. A model that performs well on a public benchmark may behave differently when connected to internal APIs, legacy systems, or messy data. The right question is not whether Large 4 beats Large 3 on a leaderboard, but whether it completes representative work at a lower cost per successful task.

Pricing: more context, higher rates

The most immediate operational change for existing Mistral customers is price. A comparison between Large 3 and Large 4 shows a substantial increase in context capacity paired with higher API rates.

• Total parameters: 675 billion for Large 3 versus 1.05 trillion for Large 4, about 56% higher.
• Active parameters per token: 41 billion for Large 3 versus 49 billion for Large 4, about 20% higher. Mistral's launch announcement cites 49 billion active parameters; the model card lists 52 billion, which would imply an increase of about 27% rather than 20% over Large 3.
• Maximum context: 256,000 tokens for Large 3 versus 1 million tokens for Large 4, about 3.9 times larger.
• Regular input price per 1M tokens: $0.50 for Large 3 versus $1.36 for Large 4, about 172% higher.
• Regular output price per 1M tokens: $1.50 for Large 3 versus $4.18 for Large 4, about 179% higher.

Large 4 is a mixture-of-experts model, which means it uses only part of its full parameter pool for each token it processes. Active parameters indicate how much of that network participates at each step. Large 4 therefore grows substantially in total size without a matching increase in active parameters, although parameter counts alone cannot establish speed, memory use, or serving cost.

Context capacity rises from 256,000 tokens to 1 million. A team could send a longer contract or a larger code repository in one request before reaching the limit. Bigger prompts can also mean more billable input tokens, so access to a larger context window can increase spending when teams use it heavily.

A two-week launch discount temporarily lowers Large 4 pricing to $0.68 per million input tokens and $2.09 per million output tokens. Even during that discount, the input price is 36% higher and the output price about 39% higher than Large 3. The discount softens the increase but does not eliminate it.

How enterprises should evaluate the upgrade

Teams considering an enterprise model upgrade should run the same representative task through both versions, then compare completion rate and final API cost with token use and latency. Cost per successfully completed task is a better measure than sticker price alone. Fewer retries or fewer calls could offset the higher rates, but that needs to be demonstrated on representative company workloads.

Buyers should also consider how much of the 1M-token context they actually need. Long-context requests can be powerful for contract analysis, large code repositories, and multi-document research, but they can also drive up input token consumption. A workflow that sends a large repository on every request may become more expensive even if the model needs fewer attempts to finish the job. Teams can mitigate this by using retrieval, caching, or summarization where appropriate, and by reserving the full context window for tasks that truly require it.

On the control side, the planned downloadable weights could change the calculus for regulated industries. Self-hosting can satisfy data residency and security requirements that a hosted API cannot. But self-hosting also means owning the model lifecycle: hardware procurement, inference optimization, uptime, patching, and compliance. For many enterprises, a hybrid approach may be more realistic, using the hosted API for experimentation and less sensitive workloads while planning private deployment for the most sensitive use cases.

Competitive context and what comes next

Mistral's move reflects a broader shift in the enterprise AI market. Model makers are competing on context length, agentic capability, and deployment flexibility, not just raw benchmark scores. Larger context windows allow models to process more information in a single request, which is useful for codebases, legal documents, incident reports, and customer histories. At the same time, pricing is becoming more complex. Enterprises must weigh input and output token costs, latency, retry rates, and infrastructure overhead.

Mistral Large 4 enters a crowded field of flagship models from multiple vendors, each with different strengths in reasoning, coding, multilingual performance, and tool use. The preview period gives enterprises a chance to test those claims before committing to a full rollout. Mistral says downloadable weights and further technical details will follow by the end of October, which will give infrastructure teams more information for deployment planning.

For now, the key facts are straightforward: Mistral Large 4 is in public preview, it offers a 1M-token context window, it posts strong but not perfect scores on multi-step enterprise benchmarks, and it costs more per token than Large 3 even with a temporary launch discount. The company plans downloadable weights and a European regional deployment. Enterprises that depend on sensitive code or proprietary data will be watching those details closely, while cost-conscious teams will compare the model's completion rate and total token spend on their own workloads. The decision will not be settled by a single benchmark or price sheet, but by whether Large 4 can complete real work reliably enough to justify its higher rates.

Source: eWeek News