Mistral Le Chonk reaches parity with GPT-6.1 Terra in coding scores

A new coding benchmark

Mistral AI recently released the open weights for its Mistral Large 4 model, which the company has nicknamed Le Chonk. According to reports from shattered.io, the model achieved a score of 48 on the Artificial Analysis Coding Agent Index. This specific score places it in the same bracket as OpenAI's GPT-6.1 Terra and Codex.

The Coding Agent Index is a significant metric because it evaluates agentic work rather than simple code completion. This means the model is tested on multi-step tasks where it must read an entire repository, edit multiple files, and execute commands. Achieving parity with closed models on this index suggests Le Chonk can handle complex, iterative programming workflows.

While Le Chonk is a strong performer in coding, it is not a specialized tool for that purpose alone. Mistral is positioning the model as a general-purpose flagship capable of handling manufacturing, finance, and multimodal tasks. However, its ability to handle complex coding loops makes it a primary candidate for developers building autonomous software agents.

Efficiency and rapid iteration

One of the most notable aspects of the Le Chonk release is its token efficiency compared to other open-weight models. Mistral claims the model uses 47% fewer tokens than the leading open-weight competitor, GLM-5.3, on the Coding Agent Index. Using fewer tokens translates to lower costs for users and faster completion times for agents that perform many sequential steps.

The model's performance improved significantly during a short training window between its public preview and its final release. On the Terminal-Bench 4.0, which measures a model's ability to drive a command line, the score jumped from 28% to 41% in less than a month. Mistral attributed these gains to continued reinforcement learning and feedback from the preview period.

Developers can also choose between low, medium, and high reasoning modes to balance accuracy against spending. By selecting a lower reasoning mode for simpler tasks, teams can further optimize their costs. These controls allow for a more granular approach to managing the trade-off between the model's intelligence and the price of the output.

Licensing and data sovereignty

Mistral released Le Chonk under the Mistral Ecosystem Open-Weight License, which functions similarly to a modified MIT license. This allows most companies to download the weights, fine-tune the model on private data, and commercialize the results. This is a major advantage for enterprises that need to keep their sensitive source code inside a private firewall.

There is a specific catch for cloud providers and resellers who wish to offer the model as a service. These entities must obtain a separate commercial license from Mistral rather than using the standard open-weight terms. This distinction ensures that while startups can build products using the weights, third-party API hosts must still coordinate with Mistral directly.

For many regulated industries, the ability to host the model on local hardware is a primary selling point. Because Mistral trained the model on Nvidia Grace Blackwell GPUs in European data centers, it appeals to regions seeking technological sovereignty. This setup allows teams to audit the model's behavior and pin specific versions to ensure consistent performance over time.

The broader intelligence landscape

Despite its success in coding, Le Chonk does not lead the field in general intelligence. On the Artificial Analysis Intelligence Index, which covers a wider variety of tasks, the model scored 26. This puts it behind several other models, including DeepSeek V4.1 Flash and GLM-5.3, which both scored 32.

The difference in these scores highlights that Mistral prioritized specific capabilities like coding and cybersecurity over broad general knowledge. The company's strategy appears to be creating a model that excels in high-value professional domains rather than trying to win every general-purpose benchmark. Buyers should weigh these different indexes based on their specific use cases.

The release also signals a shift in the competitive landscape for closed-source providers. By offering a model that reaches parity with Terra on coding tasks at a lower price point, Mistral puts pressure on vendors to justify their premiums. Tool builders may find it increasingly attractive to offer Le Chonk as a cheaper, high-quality default for their own applications.

What comes next for developers

The rapid improvement seen between the preview and the final weights release suggests that Mistral's iterative loop is highly effective. The company released the weights just 23 days after a public preview, using that time to refine the model based on real-world feedback. This cycle of rapid deployment and refinement may become a standard for how open-weights models are developed moving forward.

For teams currently using closed APIs, the next step is to evaluate the cost-benefit of switching to a self-hosted model. Mistral's pricing of $0.70 per million input tokens and $2.10 per million output tokens provides a clear baseline. Furthermore, the 90% discount on cached input tokens makes it particularly efficient for agents that frequently reference the same large codebases.

Because public benchmarks can vary, the best way to verify Le Chonk's utility is through internal testing. Developers should run a set of 20 to 50 real-world tasks from their own backlog to compare success rates. This will help determine if the model's high scores on public indexes translate to success on messy, real-world internal repositories.