Anthropic launched two models on Tuesday, Claude Fable 5 for general use and Claude Mythos 5 for a limited group of cyberdefenders and infrastructure providers, and the announcement represents the largest capability jump in Claude's history.
Fable 5 exceeds the capabilities of every model Anthropic has previously made available to the public. It is state-of-the-art on nearly all tested benchmarks across software engineering, knowledge work, vision, scientific research and tasks that require extended autonomous work.
It costs $10 per million input tokens and $50 per million output tokens, less than half the price of the previous frontier model, Claude Mythos Preview.
The Pokemon FireRed detail, tucked into the vision section of the announcement, is the one that captures something about what has actually changed.
Previous Claude models could play Pokemon FireRed, but only with a complex helper harness, additional tools that provided maps, navigation aids and extra game-state information to compensate for the limits of what the model could perceive and plan from screenshots alone. Fable 5 beat the game with a minimal, vision-only harness.
It looked at raw game screenshots and figured out the rest. The model completed something its predecessors needed scaffolding to attempt, using only its eyes.
That is not a trivial upgrade.
What Fable 5 Can Do That Previous Models Could Not
The capabilities Anthropic is claiming for Fable 5 span every dimension that matters for practical AI work, but the software engineering results are the ones most likely to land hardest for the developers and companies paying attention.
Stripe gave Anthropic the most striking data point. Fable 5 performed a codebase-wide migration in a 50 million line Ruby codebase in a single day. Stripe reported that the same migration would have taken a whole team more than two months to complete by hand.
Months of engineering compressed into a day, in a codebase of 50 million lines. The model did not just write code, it understood the codebase, identified what needed to change across its entire scope and executed the migration at a scale and speed that human engineering teams cannot match.
On Cognition's FrontierCode evaluation, which tests whether models can pass difficult coding tasks while meeting the standards of high-quality production codebases, Fable 5 scores highest among frontier models, even at medium effort.
The token efficiency matters there: a model that accomplishes more with fewer output tokens is cheaper to use and faster to run, which are the practical properties that determine whether AI coding assistance is something developers use for everything or something they reserve for specific high-value tasks.
The knowledge work results are similarly substantial. On Hebbia's Finance Benchmark for senior-level reasoning, an evaluation designed to test whether a model can do the kind of analytical work that a senior financial analyst would do, Fable 5 scored highest of any model tested. IMC, the trading firm, tested Fable 5 on trading-analysis evaluations covering factual lookup, conceptual reasoning, root-cause analysis and expected-value analysis and reported it aced them nearly across the board.
The Vision Capabilities That Are New
Fable 5 is the new state-of-the-art for AI vision. The Pokemon FireRed example is the most vivid, but the practical applications are in tasks that require understanding complex visual information, scientific figures with precise numerical data, UI screenshots that need to be reverse-engineered into code, visual layouts that previous models could describe but not act on with the precision that a task like rebuilding a web app from its screenshot requires.
The vision-only Pokemon success matters because it is a test of integrated capability, a model that can plan, navigate, remember state and execute decisions from raw visual input, in a game environment that was not designed to be AI-accessible, without the helper tools that made previous AI Pokemon runs possible.
If you can do that, the range of vision-based professional tasks you can handle expands considerably.
The Autonomous Work That Runs For Days
One of the more technically significant claims in the announcement is the long-context memory result. Fable 5 stays focused across millions of tokens in long-running tasks and improves its outputs using its own notes, a capability that matters for the class of tasks where the work takes days rather than minutes.
The Slay the Spire test quantified this, with persistent file-based memory, Fable 5's performance improved three times more than Opus 4.8's under the same conditions, and it reached the game's final act three times more often.
The agentic demonstrations included with the announcement are more visceral. Fable 5 built a solar system simulation from physics first principles and used it to predict solar eclipses.
It played Factorio, the factory-building game whose complexity is beloved by engineers specifically because optimizing a factory requires systematic thinking across many interacting systems, and built an automated factory autonomously.
It designed a 3D-printable model using a browser-based CAD editor that Fable 5 itself had also created, including an AI copilot within the editor. It coded a fluid simulation with motion synchronized to music, which Fable 5 composed from code, having never heard music.
These are not demonstrations of answering questions. They are demonstrations of an AI that builds things, designs things, and creates things over extended autonomous work sessions.
Mythos 5 And What It Does For Science
Claude Mythos 5 is the same underlying model as Fable 5 with certain safeguards lifted for a specific group of trusted users.
It is deployed through Project Glasswing in collaboration with the US government and has, Anthropic says, the strongest cybersecurity capabilities of any AI model in the world. The drug design and scientific research results come from Mythos 5.
Anthropic's internal protein design experts used Mythos 5 to accelerate aspects of the drug design process by approximately ten times.
With protein design and bioinformatics tools but no human assistance, the model matches or beats skilled human operators. It executes all the tasks that a scientist would perform: choosing binding sites, selecting and running protein design tools, and recovering from failures along the way.
Nine of the 14 protein targets studied yielded strong candidates for drug design that Anthropic is currently investigating.
Mythos 5 is also the first Claude model to consistently produce novel, compelling scientific hypotheses.
In blinded comparisons, Anthropic's scientists preferred Mythos's molecular biology hypotheses over Opus-class models roughly 80 percent of the time. One hypothesis — a novel mechanism for an E. coli protein, was independently corroborated by a separate lab that was working on the same problem without knowledge of Mythos's work.
The genomics research result is the one that attracted attention from the scientific community on Tuesday.
Mythos 5 assembled single-cell data for millions of cells spanning 138 animal species, designed and trained a custom machine learning model to identify cells performing the same role across distantly related species, and produced a trained model that outperformed a recently published study in the journal Science. The Mythos model is 100 times smaller than the one it outperformed.
The Safety Caveat And What It Means
Releasing a model of Fable 5's capability to the general public required a specific safety tradeoff that Anthropic described with unusual transparency.
Fable 5's capabilities in areas like cybersecurity could, without safeguards, be misused to cause serious damage. The solution is a system where certain topic queries are automatically redirected to Claude Opus 4.8 instead of Fable 5.
The safeguards are tuned conservatively, Anthropic expects them to trigger in less than 5 percent of sessions on average, and they will sometimes catch requests that are entirely harmless.
Anthropic said it will work to improve the safeguards and reduce false positives as more capable models arrive in the coming months.
The transparency about the tradeoff, publishing that they've conservatively tuned a system that will sometimes incorrectly restrict harmless requests, rather than pretending the system is perfectly calibrated, reflects how Anthropic communicates about the difficulty of the safety problem.
Fable 5 is available today at claude.ai and via API.