Anthropic announced Claude Fable 5 and Claude Mythos 5. Fable 5 is a Mythos‑class model released for general use with safety constraints; Anthropic says its capabilities exceed those of any previously broadly available model. Mythos 5 uses the same underlying model but has some safeguards relaxed in specific areas (notably cyber capabilities) and is available only to vetted, trusted partners.
The initial launch date was June 9, 2026. Anthropic suspended access on June 12, apologizing for the disruption, and redeployed the models on July 1, 2026.
Capabilities
Anthropic reports that Fable 5 achieves state‑of‑the‑art results on many capability benchmarks and shows especially large gains on longer, more complex tasks across software engineering, knowledge work, vision, memory, and life‑science research. Representative claims from the announcement include:
- Software engineering: Stripe reported that Fable 5 compressed months of engineering work into days; on a 50‑million‑line Ruby codebase it performed a codebase‑wide migration in a day that would otherwise have taken a team more than two months. On Cognition’s FrontierCode evaluation Fable 5 scored highest among frontier models.
- Knowledge work: On Hebbia’s Finance Benchmark for senior‑level reasoning, Fable 5 achieved the top score. IMC reported strong performance on trading‑analysis evaluations across factual lookup, conceptual reasoning, root‑cause analysis, and expected‑value analysis.
- Vision: Fable 5 can extract precise numbers from detailed scientific figures and reconstruct web app source code from screenshots. It also required less scaffolding to complete vision‑heavy tasks, for example finishing Pokémon FireRed with a minimal vision‑only harness.
- Memory and long context: Fable 5 maintains focus across millions of tokens and benefits substantially from persistent, file‑based memory in long tasks.
- Life sciences and protein design: Using Mythos 5, internal protein design experts reported roughly a 10× speedup in parts of the drug design process; in one internal study nine of 14 protein targets yielded strong candidates. Mythos 5 also consistently produced novel molecular biology hypotheses that scientists preferred about 80% of the time versus Opus‑class models, and some hypotheses advanced to experimental evaluation.
Safety approach and restrictions
Because Mythos‑class capabilities could uplift malicious actors, Anthropic deployed conservative, broad safety classifiers for Fable 5. When classifiers detect requests related to certain sensitive areas, responses are automatically handled by Claude Opus 4.8 instead; users are informed when a fallback occurs. Anthropic reports that classifiers trigger in under 5% of sessions on average and that over 95% of Fable sessions have no fallback.
The classifiers cover three main domains:
-
Cybersecurity: Mythos‑class models excel at finding and exploiting software vulnerabilities and can support agentic offensive cyber operations. Anthropic’s cybersecurity classifiers are designed to block both exploitation and broader offensive cyber tasks. Internal and external red‑teaming, including a >1,000‑hour external bug bounty, did not find universal jailbreaks; the company acknowledges preventing universal jailbreaks entirely may be impossible but aims to make any remaining exploits slow and detectable.
-
Biology and chemistry: Because well‑resourced malicious actors might seek uplift and models can now carry out practical scientific tasks, Anthropic expanded biological safeguards beyond a narrow prior list. For now, many biology and chemistry requests to Fable 5 fall back to Opus 4.8. Anthropic intends to open a trusted access program for certain biomedical researchers to use Mythos‑class biology capabilities with reduced safeguards.
-
Distillation: Attempts to extract or “distill” the model’s capabilities for training competing models (a risk observed at scale) are flagged and routed to Opus 4.8.
Anthropic says the classifiers are intentionally cautious and will be refined to reduce false positives over time.
Data retention and logging
Traffic on Mythos‑class models will be retained for 30 days on both first‑party and third‑party surfaces. Anthropic states this data will not be used to train future Claude models or for non‑safety purposes, and that human access to the data will be logged; in almost all cases data will be deleted after 30 days. The retained data will be used to detect complex or novel attacks (including new jailbreaks) and to reduce false positives.
Claude Mythos 5 and trusted access
From the redeployment, users who had access to Claude Mythos Preview (for example, cyber partners in Project Glasswing) can upgrade to Claude Mythos 5. Anthropic plans, in consultation with the US government, to add partners periodically and to open a broader trusted access program for cybersecurity organizations. A separate trusted access program for biology will enroll a small number of life‑science researchers and companies; in that program Fable 5’s biology and chemistry safeguards may be lifted while cyber safeguards remain.
Availability and pricing
Claude Fable 5 is globally available. Claude Mythos 5 is restricted initially to Glasswing partners (with cyber safeguards lifted) and to a select group of biology researchers (with biology and chemistry safeguards lifted) until the trusted access program expands. Pricing for both models is $10 per million input tokens and $50 per million output tokens. Developers can access claude‑fable‑5 via the Claude API.
Anthropic expects high and hard‑to‑predict demand. On the Claude API and consumption‑based Enterprise plans, Fable 5 is fully available from the launch date. For subscription plans the rollout is staged: through June 22, 2026, Fable 5 was included on Pro, Max, Team, and seat‑based Enterprise plans at no extra cost; from June 23, 2026, using Fable 5 on those subscriptions requires usage credits unless capacity allows extensions. Anthropic intends to restore Fable 5 to subscription plans when capacity permits and will communicate changes in advance.
Summary
Anthropic released two variants of the same Mythos‑class model: Claude Fable 5 for broad use with conservative classifiers that fallback to Claude Opus 4.8 on sensitive topics, and Claude Mythos 5 with relaxed safeguards for vetted cyberdefense and select biomedical partners. The models show substantial capability improvements across many domains, but Anthropic emphasizes cautious deployment, classifier‑driven restrictions on sensitive queries, 30‑day data retention for safety purposes, and a phased access expansion via trusted access programs.
