MISTRAL MEDIUM 3.5
Released April 2026 (v26.04). Optimized for agentic and coding workloads. 128K context. Native vision. $0.40/$2.00 per million tokens.
Mistral Medium 3.5 (released April 2026 as v26.04) is the mid-tier model in Mistral's 2026 lineup, deliberately tuned for agentic workflows and coding. It runs a 128K context window, accepts image input, and is priced at $0.40 per million input tokens and $2.00 per million output — about half the cost of Mistral Large 3 and roughly 1/12 of GPT-5.5 input. On benchmarks it lands close to Large 3 on coding while shedding the heaviest reasoning workloads. The case for: high-throughput agent loops where you want frontier-adjacent quality but cannot stomach GPT-5.5 economics, code-generation pipelines that don't need SWE-bench leadership, and multimodal classification at scale. The case against: extended reasoning chains (Large 3 or Opus 4.7 still win), and anything that needs more than 128K of context. Web search is disabled in Council AI for Mistral models.
Pick Large 3 for the hardest reasoning, longest context (256K), and where the extra ~$0.40/M input is worth it. Pick Medium 3.5 for agent loops, code generation, and bulk multimodal — anywhere throughput and cost matter more than the last few benchmark points.
April 2026 release; Mistral rebuilt the tuning specifically around agentic tool-use and code generation. The model card emphasizes function-calling reliability and lower hallucination on tool arguments.
Both are mid-tier 2026 models. Medium 3.5 is cheaper on input, has vision, and is tuned harder for agentic work. GPT-5.4 Mini has stronger tool-calling stability and native web search through the Responses API.
No — it is the active production Medium model in Council AI as of mid-2026. (Note: Mistral's prior Magistral Medium / Small via OpenRouter were deprecated separately.)
$0.40 per million input tokens and $2.00 per million output. Council AI bundles it inside monthly plan budgets — no separate Mistral key required.
Yes — pair it with GPT-5.5 for tool reliability and Claude Sonnet 4.6 for code quality. The cheap Medium 3.5 input price keeps total council cost low.