Claude Fable 5 generated a pure C++ Vampire Survivors-style demo, Lentils says

Why it matters

The demo turns Anthropic's agentic-coding pitch into a concrete artifact: a multi-file C++ game project that claims to generate its own assets at runtime, with only two user prompts.

Lentils (@Lentils80) posted a Claude Fable 5-generated Vampire Survivors-style game on Saturday, saying Anthropic's newest model produced a pure C++ project with no engine and no asset files, with sprites, sounds, graphics and music generated by code at runtime.

https://x.com/Lentils80/status/2073469967308030384

The demo, shown in a four-post thread on X, is small by game-industry standards and large by agent-demo standards. Lentils says the game came from two user prompts: one to make the game, and a second to improve it, while running in an "agentic harness." The claim matters because the project is aimed at a constraint that still exposes most AI coding demos: models can produce snippets and app shells quickly, but games test state management, rendering, animation, collision, audio and asset production in one loop.

The public source archive Lentils linked gives the demo more substance than a video clip alone. The Gofile folder is named "vesperfall-fable1914" and shows a build.bat file alongside a src directory. Inside src, the listing shows 19 files, including assets_sprites.cpp, assets.cpp, audio.cpp, entities.cpp, game.cpp, renderer.cpp, ui.cpp and weapons.cpp. That file layout is consistent with the claim that the project is carrying its visuals, audio and gameplay systems in source code, although the archive listing alone does not prove that every output was generated by the model or that the game builds cleanly on other machines.

The timing helped the clip travel. Anthropic restored Claude Fable 5 access this week after a short suspension tied to U.S. export controls. In a June 30 post, Anthropic said Fable 5 and Mythos 5 were released on June 9, suspended on June 12 after a government directive, and restored after the controls were lifted. The same post said Fable 5 would return globally starting July 1 on Claude Platform, Claude.ai, Claude Code and Claude Cowork.

Anthropic has positioned Fable 5 as its model for long-running agent work. On its Claude Fable 5 product page, Anthropic says the model is built for "hardest knowledge work and coding problems," can run in an agent harness like Claude Code or Claude Managed Agents, and is priced at $10 per million input tokens and $50 per million output tokens. The company also says Fable 5 uses safeguards for cybersecurity and biology, with some requests routed to Claude Opus 4.8, and that Fable usage requires 30-day data retention for safety monitoring.

Those constraints make Lentils' demo a useful test case for where Anthropic wants Fable 5 to sit: expensive enough that casual use has a cost, guarded enough that sensitive requests may be rerouted, and capable enough that users are pushing it into multi-file code generation rather than chat-window examples. The prompt count, if accurately described, is the sharper part of the claim. A two-prompt build compresses a project plan, implementation pass and improvement pass into a workflow that looks closer to delegation than pair programming.

The demo also shows why "no asset files" has become a meaningful benchmark in AI-generated software. In normal game production, even a clone depends on external art, sound effects and music loops. Asking a model to synthesize those elements procedurally forces the code to carry the whole experience. It reduces the legal and packaging mess around downloaded assets, while making the output easier to inspect as a self-contained program. It also shifts the weakness from asset hunting to taste, performance and maintainability.

The unanswered parts are the important ones. Lentils has not published a full prompt transcript in the thread, so the amount of hidden scaffolding inside the harness is unclear. The Gofile listing shows a Windows-oriented build file and C++ source files, but it does not establish portability, stability, licensing terms or how much manual intervention happened between the model's output and the posted video. A polished clip can hide broken edges in menus, enemy scaling, audio timing and collision behavior.

Still, the file list separates this from the usual AI coding stunt. weapons.cpp, entities.cpp, game.cpp and ui.cpp suggest a model-produced structure that at least tries to break the project into game systems rather than dumping everything into one file. audio.cpp and the asset files point to the harder part of Lentils' claim: generating the sensory layer in code instead of loading files from disk.

For builders, the practical takeaway is narrow and useful. Claude Fable 5 is being tested less like a chatbot and more like a temporary junior engine programmer with a code generator's tolerance for tedious output. The economics are different from running a free model for one-off snippets, and the reliability bar is higher than a social video. But the shape of the experiment is where frontier model coding is headed: fewer prompts, larger codebases, and more work moved into agent harnesses that can plan, write, revise and package a runnable artifact.