Blab News
Anthropic Launches Claude Fable 5, a Publicly Available Mythos-Class AI Model with Built-In Safeguards
Anthropic unveiled its most powerful Claude model yet, Claude Fable 5, on Tuesday, making a Mythos‑class large‑language model available to the public. The company claims the new model outperforms every prior release on software‑engineering, finance, vision, and life‑science benchmarks, and it comes with a built‑in safety layer that blocks queries about cybersecurity, biology, chemistry, and distillation.
Fable 5 is delivered through the Claude API at a rate of $10 per million input tokens and $50 per million output tokens, less than half the price of the earlier Mythos preview. The model is immediately usable on all consumption‑based Enterprise plans and on the Pro, Max, Team, and seat‑based Enterprise subscription plans until June 22. After that date, usage will require credits, and Anthropic says it will reinstate Fable 5 as a standard feature of subscription plans as capacity permits.
The launch follows a limited preview of Claude Mythos, which was offered to a small group of cyber‑defenders and critical‑infrastructure providers under Project Glasswing. Mythos 5 shares the same underlying architecture as Fable 5 but omits the safety safeguards that block high‑risk queries. Access to Mythos 5 remains restricted to Glasswing partners, with plans to expand a trusted‑access program that will bring additional cybersecurity and, later, biology researchers into the fold.
Early tests underscore Fable 5’s capabilities. In a collaboration with Stripe, the model completed a code‑base‑wide migration of a 50‑million‑line Ruby project in a single day—a task that would normally require a two‑month effort from a team. On the Cognition FrontierCode benchmark, Fable 5 topped the frontier models even at medium effort. In finance, it earned the highest score on Hebbia’s Finance Benchmark for senior‑level reasoning and performed strongly on IMC’s trading‑analysis evaluations.
Vision benchmarks demonstrate that Fable 5 can extract precise numbers from scientific figures and reconstruct web‑app source code from screenshots. The model also needs less scaffolding than earlier Claude releases; for example, it beat the Pokémon FireRed game with a minimal vision‑only harness.
In life‑science research, Anthropic’s internal protein‑design team used Mythos 5 to accelerate drug‑design tasks by roughly tenfold. In one study, the model matched or outperformed skilled human operators in selecting binding sites, running protein‑design tools, and recovering from failures. Mythos 5 also produced novel molecular‑biology hypotheses that researchers preferred in 80 % of head‑to‑head comparisons with Opus‑class models, and one hypothesis was later confirmed by an independent laboratory.
The company emphasized that deploying a model with such advanced capabilities carries risks. To mitigate misuse, Fable 5 is protected by safety classifiers that detect requests related to cybersecurity, biology, chemistry, or distillation. When a classifier triggers, the model automatically falls back to Claude Opus 4.8—a highly capable but less powerful model. Anthropic says more than 95 % of Fable 5 sessions do not trigger a fallback, and the classifiers have been tested against over 1,000 hours of external red‑team testing with no universal jailbreaks found.
In addition to the classifiers, Anthropic has implemented a 30‑day data‑retention policy for all traffic on Mythos‑class models. The company will not use the data to train new Claude models or for non‑safety purposes, and it will delete the data after 30 days in most cases.
Project Glasswing, which began in April, has already helped cyber‑defenders identify and fix critical software vulnerabilities. Anthropic says it will continue to expand access to Mythos 5 through a trusted‑access program for cybersecurity organizations and, later, for biology researchers.
The public release of Claude Fable 5 marks a significant step in Anthropic’s strategy to make advanced AI capabilities available while maintaining safety controls. The company says it will keep refining safeguards and reducing false positives as new models arrive.
Today, Fable 5 is available to all users under the stated pricing and usage conditions, while Mythos 5 remains limited to vetted partners. Anthropic plans to monitor usage, update classifiers, and expand trusted access programs in the coming months, and it has not announced any regulatory investigations or legal actions related to the release.
Fable 5 is delivered through the Claude API at a rate of $10 per million input tokens and $50 per million output tokens, less than half the price of the earlier Mythos preview. The model is immediately usable on all consumption‑based Enterprise plans and on the Pro, Max, Team, and seat‑based Enterprise subscription plans until June 22. After that date, usage will require credits, and Anthropic says it will reinstate Fable 5 as a standard feature of subscription plans as capacity permits.
The launch follows a limited preview of Claude Mythos, which was offered to a small group of cyber‑defenders and critical‑infrastructure providers under Project Glasswing. Mythos 5 shares the same underlying architecture as Fable 5 but omits the safety safeguards that block high‑risk queries. Access to Mythos 5 remains restricted to Glasswing partners, with plans to expand a trusted‑access program that will bring additional cybersecurity and, later, biology researchers into the fold.
Early tests underscore Fable 5’s capabilities. In a collaboration with Stripe, the model completed a code‑base‑wide migration of a 50‑million‑line Ruby project in a single day—a task that would normally require a two‑month effort from a team. On the Cognition FrontierCode benchmark, Fable 5 topped the frontier models even at medium effort. In finance, it earned the highest score on Hebbia’s Finance Benchmark for senior‑level reasoning and performed strongly on IMC’s trading‑analysis evaluations.
Vision benchmarks demonstrate that Fable 5 can extract precise numbers from scientific figures and reconstruct web‑app source code from screenshots. The model also needs less scaffolding than earlier Claude releases; for example, it beat the Pokémon FireRed game with a minimal vision‑only harness.
In life‑science research, Anthropic’s internal protein‑design team used Mythos 5 to accelerate drug‑design tasks by roughly tenfold. In one study, the model matched or outperformed skilled human operators in selecting binding sites, running protein‑design tools, and recovering from failures. Mythos 5 also produced novel molecular‑biology hypotheses that researchers preferred in 80 % of head‑to‑head comparisons with Opus‑class models, and one hypothesis was later confirmed by an independent laboratory.
The company emphasized that deploying a model with such advanced capabilities carries risks. To mitigate misuse, Fable 5 is protected by safety classifiers that detect requests related to cybersecurity, biology, chemistry, or distillation. When a classifier triggers, the model automatically falls back to Claude Opus 4.8—a highly capable but less powerful model. Anthropic says more than 95 % of Fable 5 sessions do not trigger a fallback, and the classifiers have been tested against over 1,000 hours of external red‑team testing with no universal jailbreaks found.
In addition to the classifiers, Anthropic has implemented a 30‑day data‑retention policy for all traffic on Mythos‑class models. The company will not use the data to train new Claude models or for non‑safety purposes, and it will delete the data after 30 days in most cases.
Project Glasswing, which began in April, has already helped cyber‑defenders identify and fix critical software vulnerabilities. Anthropic says it will continue to expand access to Mythos 5 through a trusted‑access program for cybersecurity organizations and, later, for biology researchers.
The public release of Claude Fable 5 marks a significant step in Anthropic’s strategy to make advanced AI capabilities available while maintaining safety controls. The company says it will keep refining safeguards and reducing false positives as new models arrive.
Today, Fable 5 is available to all users under the stated pricing and usage conditions, while Mythos 5 remains limited to vetted partners. Anthropic plans to monitor usage, update classifiers, and expand trusted access programs in the coming months, and it has not announced any regulatory investigations or legal actions related to the release.
