Atlas

Benchmarks

← All benchmarks

Vals Index v2.2

Indexes · 2026-10-19

Vals Index v2.2 (released October 19, 2026) keeps the v2.1 GDP weights and formula: finance (Finance Agent v2.1 and EMB v1.1, weight 8.0), coding (Terminal-Bench 4.0, Vibe Code Bench v1.1 and the fixed Code Migration v1.1 subset, weight 5.6), legal (Legal Research and HLAB-Vals-1.1, weight 1.2) and Tax Agent Bench at its all-pass headline (weight 0.5), divided by 15.3. Relative to v2.1 it moves Finance Agent, EMB, Code Migration and HLAB to their new versions and switches Tax Agent Bench from weighted partial credit to all-pass. Each component except Code Migration copies the standalone leaderboard result. Component runs use their respective harnesses; mixed-model fallback aggregates are excluded from single-model observations.

Top models (higher is better)

ModelScore
Gemini 4 Argon69.2
Muse Wonder Full66.8
GPT-6 Astra65.0
Muse Wonder64.9
GPT-6.1 Sol63.9
DeepSeek V4.1 Flash 101761.3
GLM-5.359.7
Claude Haiku 5.558.7
Grok 4.758.6
MiMo-V2.6-Pro57.8
Gemini 3.8 Flash56.0
MiMo-V2.6-Flash54.8
MiniMax M3.1 Flash53.8
Hy4 Preview53.2
DeepSeek V4 Pro 081353.0
Kimi K353.0
GPT-6 Luna52.3
GLM-5.3 Flash50.9
Muse Spark Open48.6
Mercury 343.7
Beam 1.0 501B A23B43.4
Inkling29.0
Inkling-Small27.6
Loading Atlas data…