Best for · Complex coding tasks, autonomous agents, long document analysis where accuracy matters more than cost.

Watch for · New tokenizer can produce up to ~35% more tokens for the same input than Opus 4.6 — the bill grows even though the rate card didn't.

Notable · SWE-bench Verified: 87.6%

Best for · Day-to-day coding, writing, and reasoning where Opus feels like overkill and Haiku feels too thin.

Watch for · Trails Opus on the hardest agentic benchmarks; for heavy tool-use workloads, test both.

Best for · Classification, routing, lightweight chat, and high-volume production pipelines.

Watch for · Reasoning drops off quickly on hard multi-step problems; not a replacement for Sonnet or Opus there.

Best for · General-purpose agents, math-heavy work, and tasks that benefit from the Thinking mode.

Watch for · Input cost doubles past the 272K-token standard window — pin context size to avoid silent surcharges.

Notable · SWE-bench Verified: ~75%

Best for · Long documents, mixed-media input, math and scientific reasoning, and anything audio- or video-shaped.

Watch for · Prompts past 200K tokens are billed at 2× the base rate; the 2M beta is opt-in.

Notable · GPQA Diamond: ~94% (Deep Think)

Best for · High-volume multimodal pipelines, transcription-shaped work, cost-sensitive chat apps.

Watch for · Not tuned for the hardest reasoning; step up to Pro when accuracy matters.

Best for · Self-hosted reasoning and code generation where cost per token dominates the decision.

Watch for · Text-first; not the right choice for vision or multimodal workloads.

API prices vary by third-party host; weights are freely downloadable.

Best for · Processing very long documents, codebases, or log files where retrieval isn't enough.

Watch for · Ten million tokens is the ceiling, not the comfortable working range — quality in the middle of the window still degrades.

Self-hostable at no API cost; numbers here are representative third-party host rates.

Best for · Multilingual work, self-hosted deployments, and anyone who wants a ladder of sizes from one vendor.

Watch for · Quality varies sharply between sizes; the advertised headline numbers belong to the flagship MoE, not the small dense checkpoints.

API prices vary by host and by size; rates shown are representative of the flagship.

Best for · Multi-agent pipelines, code-heavy agents, and cost-sensitive alternatives to Sonnet-class models.

Watch for · Ecosystem and tooling are less mature than the US frontier players; plan for integration work.

Notable · HumanEval: ~99%

Rates are Moonshot's direct API; third-party hosts vary.

Best for · EU deployments with data-residency requirements and mixed open/closed stacks.

Watch for · Sparse MoE at 675B total — latency and VRAM needs can bite on self-hosted runs despite the low token price.

Pricing on Mistral's own La Plateforme differs from some resellers; rates shown are the model-card listing.