Artificial Analysis: Claude Fable 5.2 Takes #1 on the AA-3D Interactive 3D Benchmark at 1271 Elo, 67 Points Clear, and Leads the VIALS-AA Scientific Image Reasoning Benchmark at 67%

Background and Pain Points

After Claude Fable 5.2 launched, the key question was whether Anthropic's own numbers would hold up under third-party testing. Standard Q&A and coding leaderboards rarely separate frontier models anymore. Developers want to know whether a model can build a complete, good-looking interactive artifact on its own, and whether it can reason reliably about scientific images. Artificial Analysis's AA-3D and VIALS-AA target exactly these two needs, and both are independent evaluations rather than vendor claims.

Architecture Highlights

AA-3D runs each model as an agent in an offline sandbox. Given a written brief, such as an animated Starship mission from launch through booster catch to ship landing, it builds a three.js web page within 3 hours and 250 turns, using Blender, the Rapier physics engine, Python and a headless browser to test its own work. Human panelists vote on blind pairs built from the same brief, spending at least 30 seconds with each scene, and votes become Elo ratings. VIALS-AA is AA's implementation of Handshake AI's VIALS benchmark, covering life-science images such as gel blots, plasmid maps and phylogenies.

Benchmarks

AA-3D: Fable 5.2 (Max) 1271 Elo, Sonnet 5.5 1204, Opus 5.5 1202, GPT-6 Astra 1168, Gemini 4 Argon 1147, Fable 5.1 1139. Fable 5.2's 95% confidence interval of 1246 to 1297 overlaps no other model. VIALS-AA: Fable 5.2 scores 67%, ahead of previous leader Gemini 4 Carbon (64%) and GPT-6 Astra (63%). Fable averages about 37 tool calls per task, mostly cropping and viewing images, while Astra averages about 8. Stronger Anthropic models call more tools; stronger OpenAI models call fewer. All figures are AA's independent results.

Getting Started

For 3D web pages, visualizations and game prototypes, Fable 5.2 now leads the third-party board, but AA-3D used the Max setting, which is slow and costly at $8/$40 per million tokens, so try lower effort first. For scientific or inspection tasks that require reading images, give it image cropping and code execution tools and it will check its work over many steps; if call count and latency matter more, GPT-6 Astra uses far fewer tool calls. The AA-3D Explorer lets you open every scene and compare results directly.