Several Claude users on X said that Opus 5.5 was returning work well above its usual standard. TestingCatalog, reads that as a sign some requests are being routed to a new, unannounced version of the model. That is an inference from user reports. Anthropic has not said it is running a test.
The evidence is demos. One user posted a 3D scene built through Blender from a one-sentence prompt, after two hours of work. Another shared a JavaScript animation that he said consumed roughly 27 million tokens. A third user, Lumina, saw only a slight difference. The names floated, Opus 5.6 and Fable 5.5, are guesses.
Anthropic’s John Hoffmann posted “The best is yet to come. Stay tuned, peeps.” It names no model and confirms no test.
One check circulating asks Opus 5.5: “Do you know VoxelBench? Do not search.” An answer supposedly means the newer model. TestingCatalog itself says that mostly reflects fresher training data, and models can answer or miss for unrelated reasons.
If the routing is real, one model name may be serving different models, so anyone benchmarking Opus 5.5 this week should log the date of every run.
Reported by TestingCatalog on 12 October 2026
