CALIFORNIA: Anthropic has launched Claude Fable 5, its most capable model to date and, by the company’s own reckoning, the most capable AI it has ever made available to the general public. It tops nearly every benchmark the company has tested it against, outperforms rivals in software engineering, knowledge work, vision, scientific research and complex reasoning, and comes with a built-in kill switch for the bits that could cause serious harm.
The safeguards are not a small detail. Fable 5’s cybersecurity capabilities are powerful enough that Anthropic has chosen to intercept certain queries and redirect them to its next-most-capable model, Claude Opus 4.8, rather than let Fable 5 answer freely. The company has tuned these filters conservatively, acknowledging they will sometimes catch harmless requests, but says this happens in fewer than 5 per cent of sessions on average. More capable models are coming, it adds, and the false-positive rate will improve.
For a select group of cyberdefenders and critical infrastructure providers, however, there is no such catch. Anthropic is simultaneously launching Claude Mythos 5, the same underlying model as Fable 5 but with certain safeguards lifted, exclusively through Project Glasswing, a collaboration with the US government. Mythos 5 is, the company says flatly, the most capable cybersecurity AI model in the world. Access will expand through a broader trusted-access programme in due course.
Both models are priced at $10 per million input tokens and $50 per million output tokens, less than half the cost of Claude Mythos Preview, the model they effectively supersede.
The performance claims are striking. Stripe, an early tester, reported that Fable 5 compressed months of engineering work into days: in a 50-million-line Ruby codebase, the model completed a codebase-wide migration in a single day that would otherwise have taken an entire team more than two months by hand. 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.
On knowledge work, Fable 5 achieved the highest score on Hebbia’s Finance Benchmark for senior-level reasoning, with strong gains in document-based reasoning, chart and table interpretation, and problem-solving. Trading firm IMC reported that the model aced its trading-analysis evaluations almost across the board, including factual lookup, conceptual reasoning, root-cause analysis and expected-value analysis. On vision, Fable 5 sets a new state-of-the-art, extracting precise figures from complex scientific charts and, in one test, rebuilding a web application’s source code from screenshots alone.
The memory and long-context results are equally impressive. When given access to persistent file-based memory while playing the deck-building game Slay the Spire, 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.
Mythos 5’s scientific capabilities go further still. Using the model with protein design and bioinformatics tools, Anthropic’s internal experts accelerated aspects of the drug design process by around ten times. In one test, Mythos 5 operating without human assistance matched or beat skilled human operators, autonomously choosing binding sites, selecting and running protein design tools, and recovering from failures. Nine of 14 protein targets from the study yielded strong candidates currently under investigation. In molecular biology, Mythos 5 is described as Anthropic’s first model to consistently produce novel, compelling scientific hypotheses, with Anthropic scientists preferring its output over Opus-class models roughly 80 per cent of the time in blinded comparisons.
In genomics, Mythos 5 conducted more than a week of largely autonomous novel research, assembling single-cell data for millions of cells across 138 animal species and training a custom machine-learning model that outperformed a recently published model in the journal Science, despite being 100 times smaller.
On safety, Anthropic’s automated alignment assessment found Mythos 5’s level of misaligned behaviour, including deception and cooperation with misuse, to be low and broadly comparable to Opus 4.8.
The company frames the dual launch as a step towards bringing advanced AI to as many users as possible, as quickly and as safely as it can. Fable 5 for the world. Mythos 5 for those trusted to handle what the world is not yet ready for.
The gap between those two categories may be the most consequential question in artificial intelligence right now.
