A−

Lab profile

Mistral AI

Last disclosed: July 2025 · ← Back to scorecard

Reported figures

Model named
Mistral Large 2
Energy per query
Not directly reported (per 400-token response)
Water per query
45 mL (per 400-token Le Chat response)
Carbon per query
1.14 gCO₂e (per 400-token Le Chat response)

Summary

Mistral AI commissioned an externally audited lifecycle assessment of Mistral Large 2 with Carbone 4 and ADEME, the French environment agency, published in July 2025. The LCA is the only externally audited inference disclosure currently available from a frontier lab.

Methodology

Per-query figures are reported per 400-token page of generated output, not per median prompt. The assessment covers training and inference, hardware manufacturing, and electricity mix. Audit by Carbone 4 (climate consultancy) and methodological review by ADEME.

Disclosure gaps

The functional unit (per 400-token page) is not directly comparable to Google's median-prompt figures. Newer Mistral models are not yet covered. Water figures are less detailed than energy figures.

Why this grade

External audit is the strongest credibility signal in the rubric. Slightly below A only because the functional unit complicates cross-lab comparison and coverage does not extend to current flagship models.

Sources

Request a correction →   See the rubric →