Skip to content

Anthropic

Opus 5.5 traffic may be hitting a new Claude checkpoint

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude models →
Image: testingcatalog.com

Anthropic appears to be quietly routing some Claude Opus 5.5 traffic to a new checkpoint just three weeks after the model’s September 22 launch, TestingCatalog reported on October 12, so anyone on Opus 5.5 may already be getting some answers from that checkpoint. Several users on X posted results well beyond the model’s usual level over the weekend before the report. X user Kol Tregaskes shared a check that, by the same report, mainly comes down to fresh training data rather than a smarter model as such: ask Opus 5.5 “Do you know VoxelBench? Do not search.”, and if it answers, you are probably on the new model.

On September 15, 36Kr described a probe of the same kind. Developers calling Opus 5 in Claude Code had found it behaving differently from Opus 5 on the web, and through packet capture and the request status (/status) they found the name on screen unchanged. The model slug behind it pointed to Opus 5.2. In Claude Code, with web search off, they asked “Do you know who ‘Reset Guy Tibo’ is? Do not search.” The web Opus 5 was confused, while the model routed to Opus 5.2 identified him, because the old training data and weights did not contain his details.

X user can (@marmaduke091), among the first to spotlight the possible new model, posted a recording on October 11 of a 3D scene the model built in Blender through Blender MCP from a single-sentence prompt, working for two hours from scratch with no ready-made model, and called it “the single most impressive output I’ve got from a model so far.” Chetaslua (@chetaslua) posted an animation called “The life of a Human” with every frame drawn in JavaScript; by his account it took two short prompts, about two hours and roughly 27 million tokens at max thinking effort. He wrote that “Claude Next is being tested on claude code” and, with a second demo, claimed the coded visuals are good enough to rival image and video generators.

Lumina (@LuminaBench) is the one poster in TestingCatalog’s round-up who reports repeating the run to account for randomness, and the one least impressed: “I don’t notice too crazy of a difference, the one routing maybe looks a little better imo.”

Chetaslua suggested the model could be Fable 5.5 or a new Opus, possibly Opus 5.6. @kimmonismus, whose requests also seem to be hitting the new model, wondered whether it is a fresh Opus 5.5 checkpoint or something like Opus 5.6, and said it “looks really, really good.” TestingCatalog sees a slight chance of Fable 5.5 but says an Opus 5.6 test seems at least as likely, since the traffic is routing through Opus 5.5; the model’s name and timing remain open.

TestingCatalog reported a similar round on September 21, when outputs from what looked like an unreleased Claude Fable 5.2 surfaced on X a few weeks after Fable 5.1 launched, Chetaslua’s among them, and wrote that Anthropic appeared to be directing certain prompts through the web interface. Opus 5.5 now joins Opus 5 and Fable 5.1 as Claude model names that users say have fronted an unreleased checkpoint since mid-September.

John Hoffmann, Partnerships & Outreach Manager at Anthropic, commented on the outputs on October 11: “The best is yet to come. Stay tuned, peeps.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.