A handful of Claude users on X say the chatbot has started answering some of their prompts with a model that behaves nothing like the one they normally get. None of this comes from Anthropic. It comes from screenshots, and screenshots are not an announcement.
The clearest example came from a user posting under the handle Chetaslua, who shared a ship-shooter game and a matching video demo that the model reportedly built in a single pass. In a follow-up post the same day, Chetaslua said the model produced “pure JS thats it one shot” and pointed to a separate demo: a small animated world with a walking Claude mascot that tracks the cursor across the screen. Both scenes were coded directly in JavaScript, with no external image files or plugins, according to the poster. That claim has not been independently verified, and Anthropic has not confirmed which model, if any, generated the output.
A second X user, posting as JAZII, said they ran identical prompts through the mystery Claude output and through OpenAI’s newest flagship model at high reasoning effort, then judged the Claude side “a huge step up” that held its own against the competing system. Comparisons like this are common the moment a lab is suspected of testing something new, and they carry the same weakness every time: a single user, a handful of prompts, no published benchmark, and no way for anyone else to reproduce the result.
The working theory among the people posting these outputs is that Anthropic is quietly routing some web-interface traffic to a newer version of its Fable model line, the version behind Anthropic’s recent consumer-facing releases, while the public product stays on the version released a few weeks earlier. That would make this an A/B test rather than a launch, and Anthropic has said nothing to confirm or deny it either way.
Separately, a poster using the handle lyra claimed Anthropic is running an internal build referred to by the codename “claude-wafer-eap,” and floated Opus 5.5 as the name of a coming release, with a possible announcement or launch within the week. The same posts included pricing figures said to apply to the new model, rumored at roughly $4 per million input tokens and $20 per million output tokens. None of that is confirmed. It is not a company statement, it does not appear in Anthropic’s documentation, and the exact date, name, and price could all change before anything ships.
This is the pattern that now precedes most frontier releases: outputs leak through ordinary product usage before a company says a word, and the internet treats screenshots as evidence. Sometimes that pattern is right. Anthropic’s Claude Fable line has shipped fast updates before, and a follow-on release inside a few weeks would not be unusual for the company. Sometimes the pattern is wrong, and the “leaked” model turns out to be a routing quirk, a prompt variation, or a misread screenshot.
Anthropic has not responded to the claims, and nothing here should be read as a confirmed launch, a confirmed price, or a confirmed model name. Builders who rely on Claude for production workloads should treat any sudden jump in output quality as a signal to test more prompts and log the results, not as proof that a new model is live, and should wait for Anthropic’s own announcement before changing what they ship on.
Based on reporting by TestingCatalog, published September 21, 2026.
