Claude Fable 5.2 tops new life-science image benchmark as tool use nearly doubles scores

The benchmark, which tests how well AI models read complex life-science images, found that Claude Fable 5.2 scored 67%, narrowly beating Gemini 4 Carbon at 64% and GPT-6 Astra at 63%. However, the more significant discovery was the impact of tool access, which nearly doubled average scores across all models tested.

VIALS-AA benchmark details

VIALS-AA functions as a lab practical rather than a written exam, requiring models to interpret visuals scientists encounter daily, such as gel blots and molecular structures. The benchmark includes 161 tasks across eight domains, using 174 source images from real biotech workflows. Artificial Analysis published the findings between November 6-10, 2026, with Fable 5.2 itself released on November 9, 2026.

Tool-use impact

The research revealed a strong correlation between tool usage and performance. Claude Fable 5.2 averaged about 37 tool calls per task, mostly for image analysis, while its rivals were more restrained: Haiku 5.5 averaged 21 calls, and GPT-6 Astra averaged about 8. This approach helped Fable 5.2 achieve its top position.

Cost efficiency matters

A striking result showed that Anthropic's smaller and cheaper Claude Haiku 5.5 beat the flagship Fable 5.2 when Haiku had tools and Fable did not, while also being more cost-effective. The research found that cheaper OpenAI models reached scores of 59-60% at a fraction of Claude models' costs, with per-task costs varying by as much as 99x across the field.

Broader implications

Beyond VIALS-AA, Fable 5.2 has performed strongly, topping Artificial Analysis's Intelligence Index v5.0.1 and posting 80% on HLE-Diamond and 67% on AA-Omniscience v1.1. The key takeaway is that model selection and system design are now intertwined—asking which model is smartest is less useful than determining which model with which tools at which price solves the problem best.