Atlas
Benchmarks
326 sources · 1,256 models × 2,956 benchmarks · 154,484 scoresVIALS-AA (average of 3)
Correctness averaged over three repeats per task. Artificial Analysis' own run of Handshake AI's 161-task VIALS life-sciences image set in Stirrup with Python and no network, plus an image-viewing tool for vision-capable models, three repeats per task and GPT-6 Luna (medium) grading. A repeat with no answer scores zero. Handshake's own protocol is the vials benchmark, so this stays separate.
Models
19
Top score
68.3
Leader
GPT-6 Astra
Updated
—
Resolution
—
Saturation
—
Signal
42ndpercentile
450th strongest of 779 fitted benchmarks
Signal-to-noise +5.00 for the VIALS family
Difficulty
Not estimated
Another benchmark represents this family in the fit.
19 dated models · 2 frontier modelsClosed weightsOpen weights
Data: Artificial Analysisatlas.village.dev
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | GPT-6 Astra | max · Stirrup | 68.3 | |
| 2 | Gemini 4 Carbon | high · Stirrup | 67.3 | |
| 3 | Gemini 4 Argon | high · Stirrup | 66.3 | |
| 4 | GPT-6.1 Terra | max · Stirrup | 66.0 | |
| 5 | Claude Sonnet 5.5 | max · Stirrup | 65.4 | |
| 6 | GPT-6.1 Sol | max · Stirrup | 64.8 | |
| 7 | Muse Wonder | max · Stirrup | 64.8 | |
| 8 | Claude Haiku 5.5 | max · Stirrup | 60.5 | |
| 9 | Kimi K3.1 Pro | — · Stirrup | 60.0 | |
| 10 | MiMo-V2.6-Pro | thinking · Stirrup | 58.4 | |
| 11 | Inkling 2 Small | 0.99 · Stirrup | 57.8 | |
| 12 | DeepSeek V4.1 Flash 1017 | max · Stirrup | 53.2 | |
| 13 | GLM-5.3 Flash | max · Stirrup | 53.0 | |
| 14 | Grok 4.7 | x-high · Stirrup | 52.4 | |
| 15 | Gemini 3.8 Flash | high · Stirrup | 52.0 | |
| 16 | Qwen3.8 Max (0902) | thinking · Stirrup | 50.9 | |
| 17 | MiniMax M3.1 Flash | — · Stirrup | 49.7 | |
| 18 | GPT-6 Luna | max · Stirrup | 44.1 | |
| 19 | Mistral Large 4 | high · Stirrup | 42.0 |
