Claude Fable 5.2 tops VIALS-AA science visual reasoning benchmark at 67% with heavy tool use
ArtificialAnlys · x · 2026-11-10
Artificial Analysis released VIALS-AA, its implementation of Handshake AI's VIALS benchmark for visual reasoning in life sciences (gel blots, plasmid maps, phylogenies):
- Claude Fable 5.2 (max with fallback) leads at 67%, ahead of Gemini 4 Carbon (high) at 64% and GPT-6 Astra (max) at 63%
- Fable averages 37 tool calls per task — more than any Anthropic model — while Astra averages 8, the fewest tested; OpenAI models make fewer calls as they get smarter, opposite of Anthropic's trend
- A protein-interaction task trace shows Fable cropping images, thresholding color channels, fitting circles to arcs, counting bars as intensity peaks, and mapping spokes back to the original image to verify
- Fable batches calls on 28 of 51 turns (1.75 calls/turn vs 1.12 for Sonnet 5.5)
More from Models
- Anthropic's Fable 5.2 goes live on Cola at 6x pricing, aimed at hardest tasks — oran_ge · 2026-11-10
- ChatGPT/Codex desktop app's latest update breaks the send button, user reports — rudrank · 2026-11-10
- Claude Fable 5.2 beats trendline on all three unsaturated benchmarks — emollick · 2026-11-10
- Meta's Stealth Model Crouton Alpha Revealed as Early Muse Ember, 5T Tokens Processed — alexandr_wang · 2026-11-10
- OpenAI says GPT-6.1 Terra beats Sol on visual understanding; Claude Opus 5.5 tops chart — reach_vb · 2026-11-10
- Polymarket odds: Grok 5 by end of 2026 just 11%, by end of 2027 83% — Polymarket · 2026-11-10