Understanding the Dual Launch

Anthropic is essentially releasing one powerful model under two different names and access policies .

• Claude Fable 5 (API ID: claude-fable-5): The public-facing model. Billed as "a Mythos-class model that we've made safe for general use" . It includes new safety classifiers on cybersecurity, biology/chemistry, and model distillation topics .
• Claude Mythos 5 (API ID: claude-mythos-5): The restricted-access model. It uses the same underlying model and weights as Fable 5 but with safety classifiers removed in specific areas. It is not generally available and is initially limited to approved Project Glasswing partners .

Both models are built on a 1 million token context window and can generate up to 128k output tokens per request .

Pricing: A Limited Subscription Window Before Pay-Per-Token Begins

• Input tokens: $10 per million (less than half the cost of the previous Claude Mythos Preview)
• Output tokens: $50 per million

Access to Fable 5 for subscribers comes with a hard deadline:

• June 9–22, 2026: Fable 5 is included at no extra cost in Pro, Max, Team, and seat-based Enterprise plans.
• June 23, 2026: Fable 5 is removed from subscription plans. After this date, it will require usage credits or API consumption pricing . Anthropic has stated it plans to "restore it as a standard subscription feature as soon as possible" .

After June 23, Fable 5 will remain accessible through the API, cloud platforms like AWS and Vertex AI, and consumption-based Enterprise plans . Mythos 5 will never be available through standard subscriptions; it stays locked behind Anthropic's trusted-access program .

How Fable 5's Safety Classifiers Work

Instead of simply blocking a dangerous query, Fable 5 uses a set of external AI systems (classifiers) that monitor for misuse and jailbreak attempts. When a user triggers a classifier, the request is not refused. Instead, the model silently falls back to a less capable system, Claude Opus 4.8, and the user is informed .

This fallback mechanism is active in three main areas:

1. Cybersecurity (offensive capabilities)
2. Biology and chemistry (dual-use knowledge)
3. Model distillation (preventing intellectual property theft)

The company acknowledges the system is tuned conservatively and can flag harmless requests, but states that the safeguards activate in less than 5% of sessions on average . The system underwent extensive testing, including an external bug bounty that found no universal jailbreaks in over 1,000 hours of testing . Users are not charged for prompts that trigger the fallback to Opus 4.8 before any output is generated, and prompt-cache costs for switching models are refunded .

New Data Retention Policy

A key policy change accompanies the launch. Both Claude Fable 5 and Claude Mythos 5 are designated as "Covered Models" under a new data retention policy. This means they carry a mandatory 30-day data retention period and are not available under zero-data-retention agreements, which some enterprise customers may rely on .

Mythos 5: Restricted Access for Cybersecurity and Biology

Claude Mythos 5 represents the capabilities frontier without the safety nets applied to Fable 5. Its rollout is being managed through a trusted access program:

• Project Glasswing: Mythos 5 serves as an immediate upgrade to Claude Mythos Preview for a vetted community of cyberdefenders and critical infrastructure providers across 15 countries, in collaboration with the U.S. government .
• Planned Expansion: Anthropic intends to "expand access to Mythos 5 through a broader trusted access program" that will eventually include biology researchers. This is based on the model's strong early results in drug design and molecular biology .

Benchmark Performance and Real-World Capabilities

Anthropic reports its internal benchmarks and partner testing show Fable 5 as "state-of-the-art on nearly all tested benchmarks of AI capability" . Key performance highlights include:

• Software Engineering: Achieved the top score on Cognition's FrontierCode evaluation and, at Stripe, performed a 50-million-line Ruby code migration in a single day—work estimated to otherwise take a team over two months .
• Knowledge Work: Scored highest on Hebbia's Finance Benchmark for senior-level reasoning and "aced nearly across the board" a series of trading-analysis evaluations from IMC .
• Vision: Can rebuild a web app's source code from screenshots and autonomously completed Pokémon FireRed using only a minimal, vision-based interface, a feat requiring a complex setup for older models .
• Memory & Long-Context: Uses persistent, file-based memory to improve its own outputs over time. In the game Slay the Spire, this memory capability improved its performance three times more than it did for Claude Opus 4.8 .
• Life Sciences (Mythos 5): In drug design tasks, the model accelerated parts of the process 10x, matching or beating skilled human scientists end-to-end. It also became the first Anthropic model to consistently produce novel, compelling scientific hypotheses, preferred by researchers nearly 80% of the time over previous models in blinded tests .

Understanding the Dual Launch

Anthropic is essentially releasing one powerful model under two different names and access policies .

• Claude Fable 5 (API ID: claude-fable-5): The public-facing model. Billed as "a Mythos-class model that we've made safe for general use" . It includes new safety classifiers on cybersecurity, biology/chemistry, and model distillation topics .
• Claude Mythos 5 (API ID: claude-mythos-5): The restricted-access model. It uses the same underlying model and weights as Fable 5 but with safety classifiers removed in specific areas. It is not generally available and is initially limited to approved Project Glasswing partners .

Both models are built on a 1 million token context window and can generate up to 128k output tokens per request .

Pricing: A Limited Subscription Window Before Pay-Per-Token Begins

• Input tokens: $10 per million (less than half the cost of the previous Claude Mythos Preview)
• Output tokens: $50 per million

Access to Fable 5 for subscribers comes with a hard deadline:

• June 9–22, 2026: Fable 5 is included at no extra cost in Pro, Max, Team, and seat-based Enterprise plans.
• June 23, 2026: Fable 5 is removed from subscription plans. After this date, it will require usage credits or API consumption pricing . Anthropic has stated it plans to "restore it as a standard subscription feature as soon as possible" .

After June 23, Fable 5 will remain accessible through the API, cloud platforms like AWS and Vertex AI, and consumption-based Enterprise plans . Mythos 5 will never be available through standard subscriptions; it stays locked behind Anthropic's trusted-access program .

How Fable 5's Safety Classifiers Work

Instead of simply blocking a dangerous query, Fable 5 uses a set of external AI systems (classifiers) that monitor for misuse and jailbreak attempts. When a user triggers a classifier, the request is not refused. Instead, the model silently falls back to a less capable system, Claude Opus 4.8, and the user is informed .

This fallback mechanism is active in three main areas:

1. Cybersecurity (offensive capabilities)
2. Biology and chemistry (dual-use knowledge)
3. Model distillation (preventing intellectual property theft)

The company acknowledges the system is tuned conservatively and can flag harmless requests, but states that the safeguards activate in less than 5% of sessions on average . The system underwent extensive testing, including an external bug bounty that found no universal jailbreaks in over 1,000 hours of testing . Users are not charged for prompts that trigger the fallback to Opus 4.8 before any output is generated, and prompt-cache costs for switching models are refunded .

New Data Retention Policy

A key policy change accompanies the launch. Both Claude Fable 5 and Claude Mythos 5 are designated as "Covered Models" under a new data retention policy. This means they carry a mandatory 30-day data retention period and are not available under zero-data-retention agreements, which some enterprise customers may rely on .

Mythos 5: Restricted Access for Cybersecurity and Biology

Claude Mythos 5 represents the capabilities frontier without the safety nets applied to Fable 5. Its rollout is being managed through a trusted access program:

• Project Glasswing: Mythos 5 serves as an immediate upgrade to Claude Mythos Preview for a vetted community of cyberdefenders and critical infrastructure providers across 15 countries, in collaboration with the U.S. government .
• Planned Expansion: Anthropic intends to "expand access to Mythos 5 through a broader trusted access program" that will eventually include biology researchers. This is based on the model's strong early results in drug design and molecular biology .

Benchmark Performance and Real-World Capabilities

Anthropic reports its internal benchmarks and partner testing show Fable 5 as "state-of-the-art on nearly all tested benchmarks of AI capability" . Key performance highlights include:

• Software Engineering: Achieved the top score on Cognition's FrontierCode evaluation and, at Stripe, performed a 50-million-line Ruby code migration in a single day—work estimated to otherwise take a team over two months .
• Knowledge Work: Scored highest on Hebbia's Finance Benchmark for senior-level reasoning and "aced nearly across the board" a series of trading-analysis evaluations from IMC .
• Vision: Can rebuild a web app's source code from screenshots and autonomously completed Pokémon FireRed using only a minimal, vision-based interface, a feat requiring a complex setup for older models .
• Memory & Long-Context: Uses persistent, file-based memory to improve its own outputs over time. In the game Slay the Spire, this memory capability improved its performance three times more than it did for Claude Opus 4.8 .
• Life Sciences (Mythos 5): In drug design tasks, the model accelerated parts of the process 10x, matching or beating skilled human scientists end-to-end. It also became the first Anthropic model to consistently produce novel, compelling scientific hypotheses, preferred by researchers nearly 80% of the time over previous models in blinded tests .