MISTRAL SMALL 4
Released March 16, 2026 (v26.03). Merges three previous Mistral products into one. 128K context. Native vision. $0.15/$0.45 per million tokens — the cheapest Mistral.
Mistral Small 4 (released March 16, 2026 as v26.03) consolidates three previous Mistral products — Small, Nemo, and the legacy Codestral mini — into a single SKU. It runs a 128K context window, accepts image input natively, and is priced at $0.15 per million input tokens and $0.45 per million output. That makes it one of the cheapest multimodal models in the entire 2026 frontier-adjacent lineup, beating GPT-5.4 Nano on input price while offering vision GPT-5.4 Nano doesn't. The case for: bulk classification, embeddings-pipeline judging, autocomplete-style code helpers, RAG answer drafting where you have plenty of retrieved context and just need a fluent stitch. The case against: anything requiring real reasoning depth, agent loops with complex tool graphs, or contexts beyond 128K. Web search is disabled in Council AI for Mistral models — pair Small 4 with a Gemini variant when freshness matters.
Mistral's prior small-tier SKUs — the original Small, the Nemo open-weights model, and the legacy Codestral mini — overlapped in capability and confused buyers. v26.03 consolidates all three into one model that handles general chat, vision, and code, simplifying the product surface and letting Mistral focus tuning budget on a single mid-cost target.
The hosted API is what Council AI uses. Mistral has historically released open-weight versions of small-tier models, but check Mistral's release notes for the current open-weight posture of v26.03.
Three SKUs: prior Mistral Small, Mistral Nemo, and the legacy Codestral mini. Mistral consolidated them in v26.03 (March 16, 2026).
$0.15 per million input tokens and $0.45 per million output. That's roughly 1/33 of GPT-5.5 input and 1/100 of Opus 4.7 output.
Yes. Small 4 is one of the cheapest models in the lineup — it barely dents any plan's monthly budget for conversational workloads.
For general code questions, Small 4 is fine. For dedicated coding pipelines, Codestral is the purpose-built option with a 256K context and 80+ language coverage.
Slightly more on the hardest tasks. For straightforward summarization, classification, and structured extraction the gap is small. Use Medium 3.5 when the answer matters.