Truth Check/7 October 2026

“Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse to perform the task”

The claim as it travelled: Mistral, 6 Oct 2026, about rivals' scores on a third-party index.

Mistral prints no score for the closed models; the 'near zero' is its reading of a third-party index about its rivals, so it stands as Mistral's claim, not a measured figure.

Checked Against

mistral.ai ↗Primary

Read the Check

Mistral's Large 4 is a trillion-parameter model at $1.36 per million tokens in, with the weights public by the end of October - and it does the vulnerability work closed models refuse

Mistral's own page: 52 billion active parameters (the page read 49 billion on 6 and 7 October and was changed by Mistral - corrected here 8 October), trained on 3,800 Grace Blackwell GPUs in its own European datacentres, a preview API today at $1.36 in and $4.18 out per million tokens. On a test that asks a model to reproduce a real software flaw and patch it, ML4 scores 82% - while, Mistral says, Claude Opus 5.5 and GPT-6 Astra "score near zero on the same test because they refuse to perform the task".

Every claim we have checked, in the Ledger.