Claude Fable 5.2 draws its sprites in raw JavaScript

X user Chetaslua started posting clips on September 18 of what he says is Claude Fable 5.2. Anthropic shipped Fable 5.1 on September 1, so if he’s right, the successor turned up in the wild 17 days later. Chetaslua’s first post said the answer was “routed from fable 5.1”: he had Fable 5.1 selected, and a different model replied.

Several more examples like it appeared over the weekend, TestingCatalog reports. The people who shared them said their usual standard prompts were suddenly getting answers well beyond what Claude normally produces. In the first one, a ship-shooter game, the model built both the game and its video demo. Chetaslua followed up on September 20 with two canvas scenes where small animated characters follow the mouse cursor. “Fable 5.2 is pure JS thats it one shot,” he wrote, adding that the model gave each animal its own snack without being asked. A third clip shows a Claude mascot walking around a small world to wherever you click. His caption: “no assets , no mcp , no inspiration , fully coded in JS nothing else”. Each sprite and animation is drawn in code, with no image files, plugins or outside tools.

The routing appears to happen in the Claude web interface, TestingCatalog says. It comes as Anthropic merges Cowork and chat into one experience for paid subscribers. Anthropic has done this before: Fable 5.1 went through the same kind of quiet routing, with some users getting it while Fable 5 was still selected, and it launched on September 1. This time the evidence is still just a handful of clips from a few X accounts. That’s thin for a version number, but enough to fit that pattern. If the pattern holds, the routing is a late step before release, not an early one.

Not every tester reads it that way. JAZII, who ran the same complex prompts on the new model and on GPT-6 Astra at high reasoning effort, wrote on September 18 that the next Fable model had “entered stealth testing today, so i don’t think it’s coming out anytime soon”. JAZII called it a “huge step up from current fable”. The results from the two models came out very differently, and the new Fable seemed to hold its own against Astra. That comparison is one person’s prompts, with no published method behind it.

The second track is closer to a date. lyra posted on September 20 that Anthropic is stealth testing Opus 5.5, model ID claude-opus-5-5, under the codename claude-wafer-eap. Per lyra, the version skips straight to 5.5, Anthropic has added an “imminent” note, and partners got only 24 hours of access. lyra listed two changes that would matter to anyone building on the API. Thinking would always be adaptive, meaning the model decides how much to reason, with no “off” mode. And forced tool use, the option to make the model call a particular tool, would be retired. Code that relies on either would need changes.

lyra says the release is set for Tuesday, September 22, and could come as early as Monday. TestingCatalog adds that exact versions and timelines often change.