Anthropic's Fable 5 can make weirdly fun video games with the click of a button

Overview

Anthropic released Claude Fable 5 on June 9, 2026, the first generally available model in its Mythos tier, which sits above the Opus class. The headline feature is turning a single written prompt into a working, playable browser video game when run through Claude Code, so a person with no coding background gets a finished game to play. Wharton researcher Ethan Mollick tested it hands-on and said it outperformed every other public model he had used by a wide margin. Anthropic reports the model is state of the art on nearly every benchmark it ran.

Key Takeaways

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, a tier the company places above its Opus models, and it launched on June 9, 2026.
Run through Claude Code, the model turns a single text prompt into a complete, playable browser game with no game-development work from the user.
Wharton professor Ethan Mollick generated finished games such as Snake, Strata, and Duino from one prompt each and said the model beat every public model he had tested.
Mollick reported the model ran for up to a dozen hours on multi-page specifications without stopping, far past the point where earlier models lost the thread.
Anthropic says Fable 5 is state of the art on nearly every benchmark it ran, spanning software engineering, knowledge work, vision, and scientific research.
The same skills that build games also let the model play them, navigating Factorio and finishing Pokemon FireRed using only on-screen images.

Stats & Key Facts

#Released June 9, 2026 as the first generally available Mythos-class model from Anthropic
#Ran for up to 12 hours continuously executing multi-page specifications, per Ethan Mollick
#First model to break 90 percent on Hex's core analytics benchmark, a 10-point gain over Anthropic's earlier Opus model
#Persistent file-based memory improved Slay the Spire performance 3 times more than Opus 4.8, reaching the final act 3 times as often
#More than 95 percent of Fable sessions run with no fallback to the older Opus 4.8 model
#Safety safeguards trigger on average in fewer than 5 percent of sessions

What Claude Fable 5 Is and Where It Sits in Anthropic's Lineup

Fable 5 is the first model Anthropic has released from its Mythos tier, the level above Opus.

Anthropic put its Mythos-class technology into public hands on June 9, 2026 with the launch of Claude Fable 5. Until this release, the Mythos tier was held back, and Opus models were the most capable ones the company offered to the public. Fable 5 now sits at the top of the publicly available lineup.

The model runs through Claude Code, the company's coding assistant. That setup is what lets a written request become real, running software. For non-technical readers, the takeaway is simple: this is a more capable version of the assistant that already writes and edits code, now strong enough to finish projects that used to need a team.

Turning One Prompt Into a Playable Video Game

The feature drawing attention is game creation from a single written instruction.

The headline trick is taking one written prompt and producing a working, playable browser video game. A person with no game-development background types what they want and gets something they play in their browser. There is no separate step where a developer wires the pieces together.

This matters because building even a small game normally means art, code, testing, and someone to tie it all together. Compressing that into a single instruction shows how far the gap has narrowed between describing an idea and holding a finished version of it.

The Games Ethan Mollick Built in Testing

Wharton researcher Ethan Mollick generated several finished games from one prompt each.

›Snake plays like Pac-Man with a serpent that eats apples and dies if it leaves the screen.
›Strata sends the player through a network of underground tunnels to light lanterns and earn points.
›Duino is a quiet walking game built around the poetry of Rainer Maria Rilke, with verses appearing on screen during play.
›Mollick also produced an isochronic map that shows travel times between locations, which he described as strikingly accurate and detailed.

Why This Signals a Shift for Non-Technical Builders

Projects that once needed a full team are now coming from a single instruction.

Software work that used to require a full team, including games, mapping tools, and long technical specifications, is being produced from one prompt. Mollick said Fable 5 outperformed every other public model he had used by a wide margin, and that it ran for up to a dozen hours on multi-page specifications without stopping.

The staying power is the part to watch. Earlier models tended to drift or quit partway through long, detailed jobs. A model that holds a plan across many pages and many hours starts to look less like a chat tool and more like a worker you hand a brief to.

Benchmark Results Anthropic Reported

Anthropic says the model leads on nearly every benchmark it ran.

›On Hex's core analytics benchmark, Fable 5 became the first model to break 90 percent, a 10-point gain over the earlier Opus model.
›In a Slay the Spire test, persistent file-based memory improved performance 3 times more than Opus 4.8, and the model reached the final act 3 times as often.
›Anthropic describes the model as state of the art across software engineering, knowledge work, vision, and scientific research.
›The company cited a 50-million-line code migration that normally takes about two months being compressed into days.

Playing Games, Not Just Building Them

Anthropic showed the model controlling games on its own as evidence of reasoning.

Beyond building games, Anthropic showed Fable 5 playing them without help. It navigated Factorio, a factory-building game, planning and constructing an automated factory on its own. It also finished Pokemon FireRed from start to end using only on-screen images, with no maps or navigation aids.

Earlier Claude models needed an elaborate helper setup to get through Pokemon. Doing it with vision alone is how Anthropic frames the model's sense of sequence and space, presenting it as a sign of real reasoning rather than a one-off demonstration.

Safety Behavior and Reliability

Anthropic shared figures on how often safeguards and fallbacks occur.

›Safety safeguards trigger on average in fewer than 5 percent of sessions.
›More than 95 percent of Fable sessions run with no fallback to the older Opus 4.8 model.
›Anthropic reported no universal jailbreaks across more than 1,000 hours of external bug-bounty testing.
›The company says cybersecurity classifiers block the model from making progress on offensive security tasks.

Frequently Asked Questions

What is Claude Fable 5?

It is Anthropic's first generally available Mythos-class model, released on June 9, 2026. The Mythos tier sits above the company's Opus models, making Fable 5 the most capable model Anthropic offers to the public.

How does it make a video game from a prompt?

When run through Claude Code, the model takes a single written description and produces a complete, playable browser game. The user does no coding or game-development work and plays the result in a web browser.

Who tested it and what did they say?

Wharton professor and AI researcher Ethan Mollick tested it hands-on and built games including Snake, Strata, and Duino from one prompt each. He said it outperformed every other public model he had used by a wide margin and ran for up to a dozen hours on multi-page specifications.

Can the model play games as well as build them?

Yes. Anthropic showed it navigating the factory-building game Factorio on its own and finishing Pokemon FireRed using only on-screen images, with no maps or helper tools.

How reliable and safe is it, according to Anthropic?

Anthropic reports that more than 95 percent of Fable sessions run without falling back to the older Opus 4.8 model and that safety safeguards trigger on average in fewer than 5 percent of sessions. External testing over more than 1,000 hours found no universal jailbreaks.

Claude Fable 5 moves the act of building software closer to plain description, letting a single prompt produce playable games and finished tools. For non-technical readers, it is an early look at what happens when an assistant holds a plan across hours of work and delivers something you use on your own.