Share
Anthropic’s Claude Fable 5.2 has taken first place on VIALS-AA, a new benchmark from Artificial Analysis that tests how well AI models read complex life-science images. It scored 67%, edging out Gemini 4 Carbon at 64% and GPT-6 Astra at 63%.
The headline number is not the most interesting part of the results. That honor goes to tools: giving models access to them nearly doubled average scores across the board.
One result makes the point best. Claude Haiku 5.5, Anthropic’s smaller and cheaper model, beat Claude Fable 5.2 when Haiku had tools and Fable did not. It also did so at a lower cost.
What VIALS-AA actually measures
Think of VIALS-AA as a lab practical rather than a written exam. Instead of answering trivia, models must interpret the kinds of visuals scientists stare at every day, such as gel blots and molecular structures.
The benchmark contains 161 tasks spread across eight domains. Those tasks draw on 174 source images taken from real biotech workflows.
Artificial Analysis published the findings between November 6-10, 2026. Fable 5.2 itself was released on November 9, 2026.
The tool-call gap
Fable 5.2’s winning approach looked a lot like a diligent grad student with a magnifying glass. It averaged around 37 tool calls per task, mostly for image analysis.
Its rivals were far more restrained. Haiku 5.5 averaged about 21 tool calls per task, while GPT-6 Astra averaged about 8.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The research found a strong correlation between tool usage and performance. Across the models tested, tool use nearly doubled average VIALS-AA scores.
Cheaper can mean better
The Haiku result is the one procurement teams will likely circle in red. A smaller model with tools outscored a flagship model without them, and it cost less to run.
Cost differences showed up elsewhere too. According to the research findings, cheaper OpenAI models reached scores of roughly 59-60% at a fraction of what Claude models cost.
Per-task costs across the field varied by as much as 99x.
Fable 5.2’s broader showing
VIALS-AA is not Fable 5.2’s only strong result. Shortly after release, the model moved to the top of Artificial Analysis’s Intelligence Index v5.0.1.
It also posted 80% on HLE-Diamond and 67% on AA-Omniscience v1.1.
What this means for AI buyers and builders
The clearest lesson is that model selection and system design are now tangled together. Asking which model is smartest is less useful than asking which model, with which tools, at which price, solves the problem.
There is a tradeoff worth watching, though. Around 37 tool calls per task is thorough, but each call adds time and, often, cost.
With a 99x spread in task costs, even modest price cuts from any provider could reshuffle which model looks like the best deal for scientific image work.
For now, the takeaway for anyone evaluating these systems is practical. Give the models their tools before judging them, and check the price tag before crowning a winner.
