Mistral Large 3 2512

Mistral released 2025-12-01

$0.750 per million tokens, balanced

GPT-5.6 Luna is both better and cheaper.

It scores +36.4 higher and costs 40% less ($0.450/M against $0.750/M) on this workload — and it does everything this model does.

DeepSeek V4 Flash 0731 is cheaper still (86% less) but drops no image, file input.

Our take

editorial — not a measurement

Mistral Large 3 at $0.50/$1.50 is startlingly cheap for a flagship — a third of Mistral Medium 3.5's price ($1.50/$7.50), which is the inversion you would not expect from the naming. Mistral positions Medium 3.5 as the agentic long-horizon model and Large 3 as the open-weight generalist, so the names describe lineage rather than capability ranking. Read the pricing page carefully before assuming Large is the premium option. Mistral also applies a flat 90% cached-input discount rather than per-model cache rates, which is simpler than everyone else's scheme.

Strengths

  • $0.50/$1.50 — cheapest flagship-branded model here
  • Described as open-weight by Mistral
  • Flat 90% cached-input discount, simple to model
  • 262k context, multimodal input

Weaknesses

  • Naming misleads: Medium 3.5 costs 3x more and targets harder tasks
  • No Hugging Face repo listed despite the open-weight claim
  • Licence unverified
  • No independent benchmark score captured

Reach for it when

  • EU-hosted deployments with data-residency requirements
  • Cheap general-purpose work
  • Teams wanting simple cache economics

Avoid it if

  • You need long-horizon agentic capability — Medium 3.5 is the target there
  • You need verified open-weight licensing

Sources: mistral.aiopenrouter.ai

Every price dimension

Input$0.500/M
Output$1.50/M
Cached input$0.050/M

Independent scores

Intelligence15.9
Coding20.1
Agentic5.5

Sources are listed separately rather than averaged. Across the 62 models both have scored they correlate at r = 0.806 — close agreement overall, but four models rank very differently between them, and a blended score would hide exactly those.

Capability

Context window262K
Input modestext, image, file
Tool useyes
Reasoningno

Provenance

Price sourceopenrouter.ai
Fetched2026-08-24
Quality dataverified

Mistral Large 3 2512 is Mistral’s most capable model to date, featuring a sparse mixture-of-experts architecture with 41B active parameters (675B total), and released under the Apache 2.0 license.

Markdown for LLMs