Claude Fable 5.2 builds entire games on its own, but one shooter clone cost $15,000

What's going on

On November 9, 2026, Anthropic released Claude Fable 5.2, the latest version of its most powerful model line. It isn't open source. You can use it through Anthropic's claude.ai platform, through the coding tool Claude Code, or as a paid service through the API, the programming interface for developers. Demos started piling up on launch day. AI YouTuber Matthew Berman tested the model in depth, and tech site Geeky Gadgets summarized his results. Another demo that turned heads was the browser shooter "Saltline," which Anthropic engineer Felix Rieseberg had the model generate from a single prompt. Polish outlets SilesiON and PPE were among those reporting on it.

The demos show that Fable 5.2 can work on large software projects on its own for hours or even days. That comes at a steep price, though.

Entire games built by AI

The project that drew the most attention was Berman's "Gears of War" clone. He built it in what's called Ultra Code mode, an especially thorough and expensive way of working in which the model uses a huge amount of compute. The game has just one level, but it features elaborate lighting and textures, cover behind barriers, timed reloading and full firefights. According to Berman, the project used 8 to 12 billion tokens (tokens are the chunks of text a language model reads and writes). Estimated API costs came to about $15,000. Apart from one small animation glitch, he says the level works without bugs and "is fun."

Other projects from Berman's test:

• Tumble Beans, a multiplayer game modeled on "Fall Guys" with 10 original levels and realistic physics. It took over 9 billion tokens and 71 hours to build, at an estimated cost of $13,000.
• A Call of Duty-style multiplayer shooter with blocky graphics, a Portal clone that works but looks fairly weak, the visually impressive Dragon Rider, and Starlane, an F-Zero-style racing game.
• Tidekeeper, a sequel to an earlier game in which you control a crab. All the voice acting in the games is AI-generated.
• A large 3D visualization of Tokyo with traffic, pedestrians, trains and ray-traced reflections. Ray tracing is a technique that simulates light rays physically. Every building was made from scratch. It took 6.72 billion tokens, about 70 hours and roughly $11,000.

Saltline: one prompt, eight hours, $550

Felix Rieseberg's experiment was much cheaper and better documented. He works at Anthropic and co-created Electron, a framework that powers many desktop apps. A few hours after the model launched, he posted the shooter "Saltline" on X. In his words, he wanted to test how well the model handles one-off instructions with no human feedback:

"Everything in this game is made from scratch: models, sounds, textures. About $550 at API prices, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours!"

The prompt was short and casual. Rieseberg asked for a first-person shooter with photorealistic AAA-quality graphics, but "small scale, like a tiny level." His claude.md file, where you'd normally put standing instructions for Claude Code, was empty. In his view, that means the era of prompt engineering, the craft of carefully wording instructions, is basically over. He didn't say exactly which tool or settings he used.

The game is set in a rain-soaked container port with cranes and a cargo ship. The look and mood recall recent Call of Duty games. Armed with a red-dot rifle, you have to cut three radio relays while drones ("Gulls") and robot dogs ("Hounds") attack you. Saltline runs as a public Claude artifact right in your browser on claude.ai. According to reports, you don't need a subscription to play. Subscribers can copy the game into their own account and have the model modify it.

The bill ran this high despite a single prompt because of how the model works. A model that runs for hours rereads the existing code, the plan and the tool results at every step. That's why the input tokens add up so fast. Caching made the big difference: content that's already been read gets stored, and rereading it costs just $0.20 per million tokens instead of $8. By foreign media estimates, the experiment would have cost about $3,700 without the cache. As it was, each hour of work cost about $69.

Excel, DaVinci Resolve and a 3D-printed mini tower

In Berman's test, Fable 5.2 also recreated office and creative software. Berman called a browser-based Excel clone with formulas, pivot tables and a sample sales tracker "literally perfect." It took 50 hours, 8 billion input tokens and 15 million output tokens, and cost about $11,700. A recreation of the video editor DaVinci Resolve handled waveform and video display, clip splitting and magnetic snapping. Another video editor for Windows, called Kestrel, offered keyboard shortcuts, zoom and keyframes.

Smaller tasks were much cheaper:

In the balloon simulation, you can tweak speed, caliber, lighting, gravity and balloon size live, and the physics recalculates instantly. The Blender house did have a tree sitting in the middle of the driveway. Berman printed the tower in PLA plastic on a Bambu Lab printer. It weighed under 10 grams and, according to Berman, held a considerable load before it collapsed. Recreations of real events, like the launch of the Tesla Roadster into space, were less impressive.

Benchmarks, speed and pricing

In standardized tests (benchmarks), Fable 5.2 performs strongly, according to the figures Berman cites:

According to Berman, tasks that used to take 60 to 120 minutes in Ultra Code mode now take 10 minutes at most. He says the model makes the right call nine times out of ten. In a full code review at maximum reasoning depth, it found numerous bugs, security holes and UI inconsistencies.

Pricing is $8 per million input tokens and $40 per million output tokens, 20 percent less than Fable 5.1. A Fast mode that runs 2.5 times faster costs double ($16 and $80 respectively). Berman wasn't able to test it. How much Fable usage your subscription includes and when your quota resets depends on your plan type and billing period.

Reactions and criticism

According to SilesiON, Rieseberg's post had been viewed about 428,000 times by the morning of November 10, and a gameplay clip about 655,000 times. Some users said the graphics looked AAA. One even predicted that GTA 6 would be the last big game made the traditional way. But there was pushback too. Critics argued that 445 million tokens isn't a "one-shot" but one prompt worked through over eight hours, and anyone trying to reproduce it in a quick chat will be disappointed. They also pointed out that the level has no human enemies, which would be much harder for an AI to create. Some players reported stuttering or browser crashes.

Also keep in mind that all the Saltline figures come from the creator himself, an Anthropic employee. They're rounded, and there's no log of the workflow. The "AAA" label comes from the prompt, not from an independent review. Other launch-day demos were less "pure." Anthropic's Sholto Douglas admitted he had specified the mechanics, weapons and level layout of his Portal shooter in great detail. Developer Kevin Ngo also used Tripo 3D for models and Suno v6 for music in his boat game.

Our take

Fable 5.2 pushes the limit of what an AI model can do in one go without human help, from code snippets to entire games and applications. But the most spectacular demos remain out of reach for regular users. If you spend $15,000 on a single shooter level, you're doing research or marketing, not everyday development. Even Berman thinks Opus 5.5 is more efficient and cheaper, and he calls Ultra Code mode a poor fit for most users. The smaller projects in the $25 to $100 range are more exciting. They show that sophisticated simulations and visualizations are becoming affordable. What's still unclear is how well the results hold up when others try to reproduce them independently. The benchmark scores come from Berman's summary, and many of the demos come from Anthropic employees themselves.