Will AI Replace Gamers? New Anthropic Model Beats Pokémon FireRed All on Its Own

Anthropic has published the results of internal testing of their new Claude Fable 5 model, in which the AI agent managed to fully complete Pokémon FireRed relying solely on the screen feed. The devs stressed that the model didn’t get a map, access to the game’s internal state, or any hints about where to go next. The entire visual stream (locations, menus, dialogue, battles) was fed straight in, just as if a person were sitting in front of a monitor.

In the published video, you can see Claude Fable 5 booting up Pokémon FireRed, methodically leveling up its Pokémon and working through the main story. Some actions looked a bit odd: for instance, the AI would occasionally revive weak Pokémon just to have them get instantly crushed. It’s unclear whether the model figured out that tactic on its own or supplemented what it saw with info from its training data. Still, the fact remains — the game was beaten, and without the complex helper modules that earlier experiments required.

At the same time, Anthropic tested Fable 5 in other projects. In the card-based roguelike Slay the Spire, the results were notably more modest: the AI got stuck on the third level, which the devs themselves ironically described in their blog as “beginner tier.” On the other hand, in Factorio the model managed to build an automated factory — a task that requires long-term planning, resource management, and staying focused on a goal across dozens of sequential steps.

The most noticeable catch that keeps this experiment from being seen as a prelude to replacing gamers or streamers is the massive token cost. According to Anthropic’s estimates, a full playthrough of Pokémon eats up computing resources comparable to industrial-scale volumes. Internet users have responded to the demo with a fair bit of skepticism. A comment from @nopenope-j5s spread quickly across social media.

Anthropic, of course, isn’t hiding the fact that the real goal isn’t beating Pokémon or building factories in Factorio. Games here serve as a convenient benchmark for gauging how well the model can process visual information, hold context, and adapt its strategy in a dynamic environment without human input. The skills honed in these simulations could eventually be used for UI automation, testing complex software, or managing multi‑step business processes that require end‑to‑end visual navigation. Previously, the new Anthropic model was also tried out for game creation, where it successfully made a working Minecraft clone.

What do you think — will these kinds of gaming experiments lead to real-world applications, or will they just stay flashy demos of what large models can do? Share your thoughts in the comments.