Anthropic Launches Claude Fable 5 and Mythos 5: Agentic Capability Leap and Tiered Access

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026. [1] Fable 5 is publicly available — a Mythos-class model with safety classifiers blocking cybersecurity, biology, and chemistry queries, falling back to Claude Opus 4.8 in under 5% of sessions. [1] Mythos 5 is the same underlying model with those restrictions lifted, available only through Project Glasswing to vetted US government partners and select biomedical researchers. [1][6] Both are priced at $10 per million input tokens and $50 per million output tokens — less than half the price of Mythos Preview — and the launch retires the 'Opus' branding in favor of the Fable/Mythos naming tier. [1][7][5]

For the first time, Anthropic has made a model it classifies as posing substantial misuse risk publicly available while creating a parallel restricted tier for unconstrained capability — a two-tier structure that may define how frontier labs handle dual-use AI going forward. The launch also surfaces a practical tension that will persist: safety classifiers Anthropic itself acknowledges are over-cautious, and a human-AI dynamic that early testers describe as fundamentally changing what oversight means when a model runs autonomously for hours or days.

On June 9, 2026, Anthropic simultaneously launched two models: Claude Fable 5, publicly available and carrying Mythos-class capabilities with active safety restrictions, and Claude Mythos 5, the same underlying model with those restrictions removed for a narrow set of vetted users. [1] The dual release followed weeks of pre-launch signals — the model slug 'claude-mythos-5' briefly appeared in Anthropic's API tracker in early June before being removed, prediction markets placed a 94% probability on a June 9 release date, and multiple observers reported that Anthropic was retiring the long-running 'Opus' naming tier in favor of Fable and Mythos. [2][3][4][5]

The structural split between the two models is the central design decision. Fable 5 ships with classifiers that intercept queries related to cybersecurity, biology, chemistry, and model distillation, falling back to Claude Opus 4.8 in fewer than 5% of sessions; Anthropic acknowledged these classifiers are tuned conservatively and will generate false positives on benign requests. [1] Mythos 5 removes those restrictions but is accessible only through Project Glasswing — a channel including US government partners and select biomedical researchers — and requires 30-day data retention on all traffic for misuse monitoring. [1][6] Pricing landed well below earlier signals: $10 per million input tokens and $50 per million output tokens, compared to a Mythos Preview that had been roughly 2x the current price, with some early leaks suggesting pricing as high as 5x Opus. [1][7]

Anthropic's headline capability claim is a 50-million-line Ruby codebase migration completed by Fable 5 in a single day — work estimated to require a full engineering team more than two months manually. [1][8] On the Mythos side, the model conducted autonomous genomics research over more than a week without human steering, producing a trained model that outperformed a recently published Science journal paper while being 100 times smaller. [1] Ethan Mollick, among the first external testers, described Fable as a genuine capability step but raised a structural concern: the model delegates so much of the work process to autonomous sub-agents that the human role shifts from active collaborator to what he called a 'patron' who commissions and judges finished work without observing intermediate decisions. [9] He also found the cybersecurity guardrails more disruptive in practice than the official description implies, with even faint references to security topics triggering fallback to Opus 4.8. [9]

Adoption pressure was immediate: Fable 5 exhausted the $200-per-month Claude Max usage cap for some subscribers in under 30 minutes, and the model launched simultaneously on GitHub Copilot. [10][11] On the same day, Cognition released FrontierCode, a benchmark designed to stress-test frontier models on hard coding tasks; Claude Opus 4.8 scored only 13.4% — the highest result among tested models, with GPT-5.5 at 6.3% and Gemini 3.1 Pro at 4.7%. [12] Whether Fable 5 was evaluated on FrontierCode has not been reported, leaving open how much of Anthropic's stated capability improvement translates to harder external evaluations designed to resist saturation.