Anthropic releases Claude Fable 5 and Mythos 5: The most powerful model yet enters the Mythos era

⚠️ UPDATE: Anthropic announced on June 12, 2026, that it was suspending all access to Claude Fable 5 and Claude Mythos 5 in response to a U.S. Government Export Control Directive. Officials say services are being restored as soon as possible.

Anthropic announces the launch of the Claude Fable 5 – a Mythos-level model that has been retrofitted with safety rails for widespread use. Fable 5's capabilities surpass any model previously released publicly by the company, reaching industry-leading levels in many fields such as software engineering, knowledge work, visual recognition, scientific research, etc. The more complex the task and the longer the cycle, the more obvious the lead.

Releasing such a powerful model comes with risks. If there are no security protection measures, Fable 5's capabilities in areas such as network security may be maliciously exploited, causing serious harm. Therefore, Anthropic has equipped the model with safety guardrails: queries on certain topics will be responded to by the next strongest model, Claude Opus 4.8. To ensure release as quickly and safely as possible, these security guardrails are configured conservatively—sometimes blocking harmless requests, but triggering on average in less than 5% of sessions.

Aimed at a small group of network defenders and infrastructure providers, Anthropic simultaneously released Claude Mythos 5 . This model is based on the same underlying model as Fable 5, but with security restrictions removed in some areas. Mythos 5 will be deployed first through Project Glasswing, advancing in partnership with the U.S. government.

Pricing: $10 per million input tokens, $50 per million output tokens—less than half the price of Claude Mythos Preview.

Ability Assessment

Software Engineering

According to Stripe feedback, Fable 5 compressed months of projects into days. On a 50 million line Ruby codebase, this model completed a full codebase migration in one day, whereas the same work would have taken more than two months by a human team. In Cognition's FrontierCode evaluation, the Fable 5 scored the highest of all Frontier models even at medium effort settings.

Knowledge work

In Hebbia's financial benchmark for advanced reasoning capabilities, Fable 5 achieved the highest score among all models, with significant improvements in document reasoning, chart interpretation, and problem solving. IMC noted that Fable 5 received nearly perfect scores across the board in its transaction analysis evaluation.

Visual ability

Fable 5 reaches new industry-best levels in visual tasks. The model can extract precise numbers from detailed scientific diagrams and reconstruct the source code of a web application from just a screenshot. In the Pokémon FireRed game, Fable 5 was able to complete the level using only visual cues - whereas the previous model of Claude required complex aids.

Memory and long context

Fable 5 is able to stay focused during long-running tasks involving millions of tokens. In Slay the Spire, persistent file memory allowed Fable 5 to perform three times better than Opus 4.8; the model was also three times more likely to reach the final level of the game.

Drug Design

Using Mythos 5, Anthropic's in-house protein design experts accelerated multiple aspects of the drug design process by approximately 10x. With the help of protein design tools (without human intervention), Mythos 5 has capabilities that match or exceed those of skilled human operators. Nine of the 14 protein targets yielded promising candidates for drug design.

New hypotheses in molecular biology

Mythos 5 is the first model to consistently generate novel, credible scientific hypotheses. In blind comparisons, scientists favored Mythos' molecular biology hypothesis about 80 percent of the time. One of the hypotheses - about a new mechanism in the E. coli protein - has been confirmed by a research paper from an independent laboratory.

Original research in genomics

Mythos 5 autonomously completed novel genomics research in just over a week. The model assembled single-cell data from millions of cells across 138 animal species and designed to train a custom machine learning model. With only high-level human guidance, the model it trained outperformed a model recently published in Science by a factor of 100. Results are scheduled to be published in the coming months.

Fable 5 security protection

The Mythos level model poses substantial capability enhancement risks to malicious actors. Fable 5 is equipped with a Safety Classifier System - an independent AI system for detecting potential abuse. When the classifier fires, the request will be handled by Claude Opus 4.8. Over 95% of Fable 5 sessions require no rollback.

Covering three major areas:

1. Cyber Security: Prevent vulnerability exploits and automated hacking attacks. External testing found that Fable 5’s security protection was the strongest of all tested models—zero harmful requests were successfully executed.

2. Biology and Chemistry: Block high-risk biological research inquiries. Mythos 5 demonstrates the ability to predict viral coat protein properties beyond specialized protein language models—highlighting dual-use risks.

3. Distillation Protection: Prevent the large-scale extraction of model capabilities in authoritarian countries to train competing models.

Availability and Pricing

• Fable 5: Online today, $10/$50 per million tokens
• Mythos 5: Glasswing partners only
• Subscription plan: included for free until June 22, after which points are required
• API name: claude-fable-5

The release of Fable 5 and Mythos 5 marks a new phase in the AI industry. Anthropic has established a clear model layering system - Haiku → Sonnet → Opus → Mythos/Fable - Mythos as a new layer above Opus means that AI capabilities have crossed a critical threshold and are enough to produce substantial capability improvements in high-risk areas such as cybersecurity and biological research.

Just three days after the release, the U.S. government issued an export control directive suspending access. This is the first time the U.S. government has directly intervened in the public availability of an AI model, signaling that cutting-edge AI models have become geopolitical bargaining chips. For China's AI industry, this means that models at the Mythos level and above may face continued export restrictions, and the pressure on domestic large models to catch up on the upper limit of their capabilities will further increase.

The actual measured data of Fable 5 is impressive: Stripe can complete two months of work in one day, speed up the drug design process by 10 times, and surpass the "Science" paper with a 1/100 model scale - these are not incremental improvements, but a leap in the capability level. Anthropic also demonstrated dual innovations in "capability" and "safety": the safety classifier system may become a standard configuration of future cutting-edge models, rather than an option.

Key directions worthy of continued follow-up:

1. When will Fable 5 access be restored – specific conditions and timelines for export controls
2. Progress of the biological research application channel of the Mythos 5 Trusted Access Program
3. How OpenAI/Google responds to Mythos-level capability standards
4. Subsequent optimization data of safety classifier false positive rate
5. Availability in the Chinese market – the actual impact of export controls on domestic Claude users

Reference:

• Claude System Card: https://anthropic.com/claude-fable-5-mythos-5-system-card
• Claude Official entrance: https://claude.ai/