Mistral Medium 3.1

Mistral AI officially launched Mistral Medium 3.1 on August 12, 2025, positioning it as a frontier-class, proprietary multimodal model that targets enterprise workflows in direct competition with OpenAI's GPT-4o and Anthropic's Claude 3.5. The release marks a strategic shift toward high-performance multimodal reasoning without open-weight distribution, focusing on dense visual-language integration rather than sparse token-efficient approaches. The architecture supports a 128k token context window and a 128k token output limit, with a hybrid attention mechanism that prioritizes visual grounding in text generation. The page also describes a hybrid MoE design that routes vision tokens through specialized expert networks, though that specific architectural claim is not corroborated by the other supplied sources and should be treated as single-page editorial commentary.

According to Artificial Analysis, Mistral Medium 3.1 was released in August 2025 as a proprietary model with a 128k token context window, text and image input, and text output. It is a non-reasoning model that scores 9 on the Artificial Analysis Intelligence Index, placing it above average among comparable non-reasoning models whose median score is 7, and it produced 18M tokens on the Intelligence Index evaluation, which is somewhat verbose. The publisher explicitly notes that Mistral Medium 3.1 has been deprecated in favor of Mistral Medium 3.5, so only the default 10k input token workload continues to receive updated performance benchmarking, and speed and cost metrics are listed as N/A. This supersession signal is important context when evaluating the model's current relevance for new deployments.

Galileo's model hub entry for Mistral Medium 3.1 (mistral-medium-2508) confirms the August 12, 2025 release and documents its enhanced Mixture-of-Experts architecture with native multimodal text-and-image support. The page reports a 131,072-token context window and native function calling for agentic workflows, with support for continuous pretraining and full fine-tuning for enterprise adaptation. The Galileo overview reports benchmark scores including 0.98 action completion, 0.59 tool selection, and 0.82 conversation efficiency, positioning the model competitively on cost-performance for agent deployments. The model is described as targeting enterprise agent use cases with an 8X cost reduction claim relative to larger models. These third-party benchmark figures supplement Mistral's official documentation with independent evaluation context.

Mistral AI's official documentation confirms Mistral Medium 3.1 (mistral-medium-2508) as a Premier-tier frontier multimodal model released in August 2025 with a 128k token context window. The page lists supported API features including Structured Outputs, Function Calling, Predicted Outputs, Document QnA, Prefix, Batching, Agents, and Chat Completions, giving developers a clear feature surface. This is the most authoritative first-party source for the model. The same Mistral docs page marks Mistral Medium 3.1 as deprecated, with a deprecation date of May 22, 2026, and recommends Mistral Medium 3.5 for new integrations. The model is described as improving tone and performance over its predecessor while remaining multimodal. Developers planning new builds should migrate to Mistral Medium 3.5 per Mistral's official guidance.

Mistral Medium 3.1 is the enterprise-grade, multimodal successor to Mistral Medium 3, surfaced via the OpenRouter catalog page for the exact subject slug (mistralai/mistral-medium-3.1). The listing specifies text and image inputs with text output, a 131K context window, a June 2025 knowledge cutoff, and a release date For developers accessing it through gateways such as Kilo, the page reports list pricing of $0.40 per 1M input tokens and $2.00 per 1M output tokens with a $0.04/M cache read rate. Effective 30-day weighted pricing on OpenRouter averages about $0.2645/M input tokens at a 37.6% cache hit rate, single-provider (Mistral)