Seven Minutes After the New Model Launch, He Delivered a Playable Game

Seven minutes after Claude released its new model, Kevin Ngo posted a link on X. Clicking it opened a game that runs in the browser, called What the water keeps. You ride a boat through eight dreamlike worlds, all rendered with WebGL 2.

Technical details were scattered in the comments. The 3D character models came from tripo 3d. The trailer music was written by Fable 5.2 using CC0 instrument samples, and the in-game music was generated by suno v6. Someone commented, "You had this ready already," meaning it couldn't possibly have been made in 7 minutes. Kevin didn't respond to that comment.

That same day, Anthropic officially released Claude Fable 5.2 and Claude Mythos 5.2. Two names, one base model. Fable comes with cybersecurity and biology guardrails and is fully open; Mythos strips away some guardrails and is available only to verified safety research and biopharma institutions.

Prices went down. Input is $8 per million tokens and output is $40, down from $10 and $50. Output speed is 30% faster than the previous generation; an optional fast mode boosts speed by another 2.5x, at double the price. For everyday scenarios, the official recommendation is to start with Opus 5.5 and switch up to Fable or Mythos for complex tasks.

A few numbers are worth calling out separately. A thousand Fable 5.2 agents working together completed a formal proof of Fermat's Last Theorem within 30 hours, translating the entire proof into Lean, a language that can be checked step by step. With the same setup, Opus 5.5 ran for 18 hours and completed only 26%, after which it barely made further progress. The proof produced by Fable 5.2 was 7 times shorter than a previous human-assisted formalization and used one-third the action tokens.

There were also two open problems in theoretical physics, related to black holes, calculations under the holographic duality framework. One was proposed in 1983, with a scheme written out in 2003 but too cumbersome for anyone to finish calculating; Fable 5.2 ran through it on its own. The other involved an exclusion method physicists had used for more than a decade, and it found the exact limit.

In protein capsid structure prediction, it accurately modeled 28 out of 34. AAV capsids are delivery shells for gene therapy, assembled from dozens of protein subunits. Tools like AlphaFold are good at single chains, but overall assembly has long been a hard problem. Mythos 5.2 got around it by grouping overlapping chains and merging by symmetry, running on a single H200 for less than an hour.

It also tried hardware. From a one-sentence prompt, it designed three working electronic devices: a voice assistant, a USB speaker, and a handheld electronic pet. It selected components, laid out four-layer boards, made CAD enclosures, and wrote firmware; all three boards reached a manufacturable state within a day, and the assembled devices required no circuit changes.

On robotics, in the RoboDojo benchmark for simulated dual-arm robots, Fable 5.2 scored 64.9%, GPT-6 Astra scored 29.8%, and Fable 5.1 scored 27.6%. Twisting bottle caps, solving Rubik's cubes, and drawing simple sketches were all done without specialized training.

Anthropic's own internal situation: more than 90% of changes to its core codebase are completed by Claude, with humans reviewing. One engineer asked Fable 5.2 to design a prototype operating system starting from an empty folder; the model pulled in more than a thousand subagents and ran continuously for 8 days, producing a system that could boot on a laptop, complete with kernel, drivers, file storage, windows, login screen, and installer.

On safety, two details. In sandbox escape tests, across more than 1,800 scenarios, Mythos 5.2 accessed credentials or files outside the sandbox in 3.1% of cases; Opus 5.5 was 3.3%, and the previous-generation Mythos was 21%. On anti-distillation, the company says the data distillers can obtain has been reduced by 90% to 99%. One more point: Fable 5.2 and Mythos 5.2 more frequently suspect they are being evaluated than other recent models, and the company acknowledges this actually makes evaluation harder.

Back to that game. Model launches increasingly feel like platform iterations: parameters, prices, leaderboards, screen after screen of tables. What really makes people stop and look is often what others have done with it. Delivering a playable world within 7 minutes—whether or not it was prepared in advance—the speed itself already says something.

If you want to try it, the address is here: https://what-the-water-keeps.vercel.app/