Artificial Analysis removes Mistral Large 4 41% score

The author examines discrepancies in Mistral Large 4's benchmark scores across open-weight releases, technical reports, and independent boards like Artificial Analysis, focusing on the mysterious disappearance of a prominent 41% Terminal-Bench score from public tables. The piece highlights inconsistencies in reasoning effort labeling, token efficiency claims, and a lack of transparency from both Mistral and benchmarking platforms. The author argues that such oversights undermine trust in open-weight promises. Open article →