At the latest AA-3D leaderboard check on Nov. 10, Claude Fable 5.2 held the top score at 1,271 Elo among systems tested on 55 interactive 3D scene-building tasks.
Fable finished 67 points above Claude Sonnet 5.5, which recorded 1,204 Elo, and 69 ahead of Claude Opus 5.5 at 1,202. Fable's 95% confidence interval was 1,246 to 1,297. Sonnet's range was 1,176 to 1,232, while Opus' was 1,174 to 1,231.
The 55 prompts in AA-3D v1.0 have models build interactive web scenes with three.js in an offline sandbox. Reviewers choose between two outputs for each task without knowing which model made them. AA-3D turns those preferences into Elo ratings, a relative measure based on head-to-head results. The ranking applies to this 3D task set; it does not establish which model is better at unrelated kinds of work.
The benchmark's Oct. 12 launch showed Sonnet at 1,194 Elo, statistically indistinguishable from three other models from Anthropic and OpenAI. The latest listing gives no date for Fable's evaluations or for the leaderboard update, so the check date does not establish when the model was tested.
Anthropic's Nov. 9 release said Fable 5.2 was available on the company's platforms. For API access, input costs $8 per million tokens and output costs $40 per million; Anthropic said both rates were below Fable 5.1 but did not specify by how much. Anthropic also claims Fable 5.2 can return generations as much as 30% sooner than Fable 5.1; AA-3D measured no generation times.
Sam Zhang, a creative technologist at Unity, called the model “very strong on the harder game-dev tasks.”
