Claude Fable 5 Explained: Anthropic's Public Mythos-Class Model, Safeguard Fallback to Opus 4.8, Pricing and Data Retention
Anthropic released Claude Fable 5 on June 9, 2026, a public Mythos-class model whose safeguards route flagged cyber, biology and chemistry requests to Claude Opus 4.8, alongside Mythos 5 for vetted Project Glasswing partners. Pricing, vendor-reported benchmarks, the 30-day retention policy and what developers should check.
Anthropic’s Claude Fable 5, released on June 9, 2026, makes a Mythos-class model, the capability tier Anthropic had previously restricted to vetted cyber defenders, available to the public with targeted safeguards. Flagged requests about cybersecurity, biology, chemistry and model distillation are handed to the less capable Claude Opus 4.8, and Anthropic says that happens in fewer than 5% of sessions. A twin model, Claude Mythos 5, has the cyber safeguards lifted and is restricted to Project Glasswing partners. Both cost $10 per million input tokens and $50 per million output tokens, twice the price of Opus 4.8 and less than half that of the earlier Mythos Preview (Anthropic, CNBC). The headline benchmark and customer figures are vendor-reported, and a new 30-day retention requirement matters for enterprise buyers.
At a glance
| Item | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Release | June 9, 2026 | June 9, 2026, as an upgrade to Claude Mythos Preview |
| Access | Claude API and paid Claude plans | Project Glasswing partners; select biology researchers planned |
| Safeguards | Classifiers route flagged cyber, biology, chemistry and distillation requests to Opus 4.8 | Cyber safeguards lifted for vetted users |
| API price | $10 / $50 per million input / output tokens | Same |
| Data retention | 30-day retention required for Mythos-class traffic | Same |
Sources: Anthropic; The Hacker News; MacRumors.
How the safeguards work
Fable 5 and Mythos 5 are the same model with different safeguard configurations. When Fable 5’s classifiers flag a request in a high-risk domain, the response is handed to Claude Opus 4.8, which Anthropic announced in May, and the user is told the handoff happened (The Hacker News). CNBC gave the example of a question about making the toxin ricin, which would be blocked and answered safely by Opus 4.8.
Anthropic says it tuned the safeguards conservatively, so they will sometimes catch harmless requests, but that they trigger in fewer than 5% of sessions on average, and that more than 95% of Fable sessions involve no fallback, making performance in those sessions effectively the same as Mythos 5. As The Hacker News notes, the 5% figure caps total disruption, including genuine blocks, rather than measuring false positives on their own. Security researchers and teams working near those domains should expect more fallbacks than typical users.
Why Anthropic restricted Mythos in the first place
In April 2026, Anthropic released Claude Mythos Preview only to a limited group of cyber defenders and critical software providers through Project Glasswing. According to The Hacker News, Anthropic and roughly 50 partners used Mythos Preview in the programme’s first weeks to find more than ten thousand high- or critical-severity vulnerabilities in systemically important software. Anthropic says Mythos 5 is deployed through Glasswing in collaboration with the U.S. government. The concern is straightforward: a model that finds vulnerabilities quickly can help attackers as well as defenders.
Performance claims
| Claim | Figure | Status |
|---|---|---|
| SWE-Bench Pro | 80.3% for Fable 5 vs 69.2% for Opus 4.8 | Vendor-reported |
| Stripe migration | 50-million-line Ruby codebase migrated in about a day, against a roughly two-month team estimate | Customer testimonial at launch; methodology not disclosed |
| Benchmarks vs Opus 4.8 | More than 10% higher on some benchmarks | Per Anthropic’s blog, reported by CNBC |
Sources: Tech Jacks Solutions summary of vendor claims; CNBC. Anthropic’s own post says Stripe reported that Fable 5 “compressed months of engineering into days”.
These figures had not been independently reproduced at launch. SWE-Bench Pro is an established public benchmark, so it is the claim most likely to hold up, but buyers should test on their own workloads. For why benchmark numbers need caution, see our guide to reading LLM benchmarks.
Pricing and plan access
At $10 per million input tokens and $50 per million output tokens, Fable 5 costs twice as much as Opus 4.8, according to CNBC. Fable 5 was included at no extra cost on Pro, Max, Team and seat-based Enterprise plans through June 22, 2026, after which using it requires usage credits (MacRumors). One practical detail: when a request falls back to Opus 4.8, cost and quality for that request reflect Opus 4.8, so workloads that trigger frequent fallbacks will see variable results.
The 30-day retention requirement
Anthropic says it will require 30-day retention for all traffic on Mythos-class models, on both its own and third-party platforms. It says the data will not be used to train new models or for any non-safety purpose, that all human access is logged, and that data is deleted after 30 days in almost all cases (Anthropic). For enterprises with zero-retention agreements or strict data-residency rules, this is the most important part of the announcement, and legal and security teams should review it before routing sensitive workloads to Fable 5.
Market context
Anthropic has been gaining ground in business AI spending; Ramp’s AI Index, which tracks corporate card and bill-pay data, has shown Anthropic ahead of OpenAI on business adoption in recent readings. CNBC reported that Anthropic said in May that its revenue run rate had reached $47 billion. Fable 5 is positioned as a premium tier above Opus rather than a replacement for cheaper models, so most teams will use it selectively for the hardest tasks.
What developers should check
- Test on your own tasks. Compare Fable 5 with Opus 4.8 on representative workloads, measuring quality, turns needed and total cost.
- Measure fallback rates. Log how often your traffic is routed to Opus 4.8, especially for security tooling, chemistry or life-sciences applications.
- Review data terms. Confirm whether the 30-day retention requirement is compatible with your contracts and regulations.
- Route selectively. Use Fable 5 for complex, high-value tasks and cheaper models for routine work.
- Apply for trusted access if needed. Security teams that need unrestricted cyber capability must go through Project Glasswing or the planned trusted access programme.
Our analysis
Fable 5 is a notable experiment in releasing a frontier model with targeted, transparent fallbacks rather than blanket refusals or full restriction. The design is sensible: most users get full capability, and the riskiest domains are handled by a weaker model with the user informed. The trade-offs are real, though. Performance claims are vendor-reported, the price is double Opus 4.8, and the 30-day retention rule may rule it out for some regulated workloads. For most teams, the right approach is to pilot it on the hardest tasks, track fallback rates and costs, and keep cheaper models for everything else.