Coding-source update — checked November 12, 2026

Artificial Analysis’s current Mistral Vibe table lists Large 4 High at 41 on Terminal-Bench 4.0 and 48 on Coding Agent Index; Medium is 26 and 35. High now matches the vendor figures quoted below. The earlier High 26/35 observation remains a historical source snapshot, not a claim about today’s table. No published run date or checkpoint establishes why the display changed, and this alone does not prove a model improvement.

Arena’s overall text board dated November 6 lists Carbon High at 1526 ±9, marked preliminary, and Argon High at 1523 ±6. Our prior October 27 snapshot recorded 1528 and 1520. Opus 5.5 High is now 1505 ±6, versus 1509.

Small movements need context

Nine tracked scores changed; Fable 5.1 Max stayed at 1504 ±6. These are community preference estimates with overlapping uncertainty, not evidence of new model checkpoints or proven capability gains.

Mistral Large 4 now has a linked row at 1420 ±10 with no effort stated. Its older, unverified 1432 is not a trend baseline. Wonder Full, DeepSeek Flash1017, Kimi K3.1 Pro and Terra 6.1 remain gaps; similarly named variants are not substitutes.

A separate coding-source correction

Artificial Analysis currently lists Mistral Large 4 High in Mistral Vibe at 26 on Terminal-Bench 4.0; its separate Coding Agent Index is 35. Mistral’s improved open-weight release post instead advertises 41% and Index 48. We correct our AA field while retaining the vendor claim separately. Without matching checkpoints and dated runs, the discrepancy does not establish a regression.

AA’s Intelligence Index remains v5.0.1. All fifteen tracked Intelligence scores were unchanged in this check.

Newer board — checked November 9, 2026

Source update November 9, 2026: the current Benchmarks page now uses the November 8 Text and Agent Arena boards. The figures in this report remain the November 6 historical snapshot; source-check dates do not redate it. Overlapping uncertainty does not establish a model upgrade or regression.

Newer board — checked November 10, 2026

Source update November 10, 2026: the current Benchmarks page uses the November 9 Text board and November 8 Agent board. The figures in this article remain its November 6 historical snapshot. Current Kimi K3.1 Pro Max is now separately verified; changing scores, overlapping uncertainty and source checks do not establish a model upgrade or regression.