Anthropic Unveils Claude Fable 5, Mythos 5 With New Pricing, Safety

Anthropic's launch of Claude Fable 5 and Mythos 5 introduces its new "Mythos-class" AI with major changes to pricing and safety. Fable 5, the first of this class available through the Claude API, offers broad access, while Mythos 5 remains exclusive to vetted partners.

This release follows a split deployment strategy, as noted by TechCrunch, which tracks the rollout: Fable 5 is a widely available model with robust safety guardrails, while its more capable sibling, Mythos 5, is reserved for trusted researchers. Anthropic's own announcement confirms this approach, detailing the rationale behind the commercial and safety strategy.

New Pricing Structure and Access Tiers

Anthropic's Claude Fable 5 is a powerful, publicly accessible AI model with built-in safety restrictions for high-risk topics. Claude Mythos 5 is the same underlying model but with fewer safety guardrails, making it more capable but restricted to a small group of vetted research partners.

Initially, Fable 5 access is bundled at no additional cost for Pro, Max, Team, and Enterprise subscribers until June 22, 2026. After this date, usage will shift to a metered model. As reported by TechCrunch and confirmed by Simon Willison's weblog, the new rates will be $10 per million input tokens and $50 per million output tokens - double the rate of Claude 4.8 Opus. These prices also apply to direct API usage of the claude-fable-5 model.

Key Details for Subscribers:
- Promotional Access: Bundled, no-cost access ends on June 22, 2026.
- Metered Billing: After the promotional period, usage will be billed against plan credits at the new API rates.
- Universal Change: This pricing schedule applies to all Team and Enterprise plan subscribers.

Pricing for Mythos 5 has not been made public. Access is strictly controlled and limited to vetted partners, including researchers from Project Glasswing and select biology labs, according to an Anthropic press note.

Safety Guardrails and Dual-Use Risk Management

Claude Fable 5 includes what Anthropic terms "selective guardrails" to mitigate risks. Queries in sensitive domains like advanced biology, chemistry, and cybersecurity will trigger an automatic refusal or a fallback to the older Claude 4.8 Opus model. TechCrunch confirms that Anthropic retains logs for 30 days to audit these safety interventions. Anthropic's system card elaborates on these risks, warning that such models could "uplift" or accelerate the work of malicious actors.

In contrast, Mythos 5 relaxes many of these safety restrictions for its approved users. Anthropic's rationale, summarized by The Hacker News, is that a wider release of the unrestricted model could equally benefit malicious actors and safety researchers. This dual-use concern is echoed by AWS, the hosting provider for both models, which states that access to the less-restricted Mythos 5 is reserved for a "small group of vetted customers".

Early Benchmarks and Performance Reports

Anthropic claims Fable 5 demonstrates superior performance on complex, long-duration tasks like multi-hour coding projects and advanced reasoning. Internal tests indicate that 95% of user sessions completed successfully without needing to fall back to an older model. Early external benchmarks support these claims:
- CursorBench: Fable 5 achieved a new state-of-the-art score of 72.9% on this coding suite, as cited by Cursor.
- SWE-Bench Pro: A developer roundup reported a score of 80.3%, a significant improvement over the 69.2% from Claude 4.8 Opus.

Anecdotal reports from early adopters reinforce these impressive metrics. TechCrunch highlights a 90% score from analytics company Hex on its most difficult internal benchmark. In another case, Stripe reportedly used Fable 5 to complete a 50-million-line Ruby codebase migration in a single day - a project initially estimated to take two months. However, reviewers urge caution, noting these are vendor-reported results that await independent replication. Furthermore, AINews observes that the model may exhibit slower latency on more demanding tasks.

Key Takeaways: Fable 5 vs. Mythos 5

In summary, Claude Fable 5 is positioned for the general developer community, with automated safety blocks on sensitive topics. Claude Mythos 5 unlocks advanced capabilities in areas like cybersecurity and biology but is gated behind strict access controls. While early benchmarks show major performance gains in coding and analytics, the post-promotional pricing will be significantly higher, and the community awaits independent verification of the reported speedups.

---

What exactly is the difference between Claude Fable 5 and Claude Mythos 5?

Claude Fable 5 is the public-facing version of Anthropic's new "Mythos-class" model that any paying API or web user can call via the name claude-fable-5. Claude Mythos 5 is the same underlying system with fewer safety guardrails and is available only to vetted partners - currently Project Glasswing members and a small pool of biology researchers. In practice, that means Mythos 5 can answer more sensitive cyber-security or biochemical questions that Fable 5 will automatically block or route to the older Opus 4.8 model.

How much will Claude Fable 5 cost when the temporary inclusion ends?

Starting June 23, 2026, Fable 5 drops out of the Pro, Max, Team and seat-based Enterprise bundles and will be billed purely by usage:

• $10 per million input tokens
• $50 per million output tokens

That is exactly 2× the Opus 4.8 price, so heavy users on subscription plans will need to load usage credits after the June 22 cutoff unless Anthropic grants an extension.

Which tasks are actually blocked or downgraded in Claude Fable 5?

Anthropic's API-level classifiers intervene on requests judged high-risk in cybersecurity, biology, chemistry, model distillation and similar dual-use domains. When such prompts are detected, the system either:

• Refuses the request outright, or
• Routes it to Claude Opus 4.8 as a safer fallback.

Early testers report that this invisible switching can reduce output quality on advanced machine-learning and biomedical queries compared with the gated Mythos 5 variant.

How long does Anthropic retain my conversation data when I use these models?

According to the published system card, all traffic is retained for 30 days to enable abuse investigations and safety tuning - even if you are on an Enterprise plan with stricter data-residency requirements. This window applies to both Fable 5 and Mythos 5 calls.

What real-world speedups are users seeing with Fable 5 on coding tasks?

Independent early benchmarks and vendor anecdotes point to sizeable gains:

• 72.9% on CursorBench (new state-of-the-art, ~8 pts above the prior best)
• 80.3% on SWE-Bench Pro vs 69.2% with Opus 4.8
• Stripe is said to have used Fable 5 to finish a 50-million-line Ruby migration in one day that previously required two months of human effort.

Bear in mind these figures come from launch-day disclosures and early tester reports, so replication is still pending.