Anthropic on Tuesday made Claude Fable 5 available to enterprise customers and paid subscribers — but the most capable version of the same underlying model, Claude Mythos 5, remains off-limits to the general public. The divide is deliberate: Fable 5 ships with safeguards that redirect certain queries to an older model, while Mythos 5 goes only to a small, vetted group of cyber defenders and infrastructure providers. For most users, the access ceiling is built into the product itself.
What Claude Fable 5 Capabilities Actually Cover — and Where They Stop
Anthropic says Fable 5 uses the same underlying model as Mythos but adds safeguards that intercept queries on high-risk topics — particularly cybersecurity and biology — and route them instead to Claude Opus 4.8. According to Anthropic, those safeguards fire in roughly 5% of sessions, meaning 95% of Fable sessions involve no fallback at all. The company acknowledges that its classifiers can trigger on harmless requests and that no classifier system can prevent every jailbreak.
To stress-test those boundaries, Anthropic ran a 1,000-hour external bug bounty; no universal jailbreaks were found. External red-teaming was also conducted by UK AISI. The full picture of what the model can and cannot do is documented in the system card and a separate risk report.
Within permitted scope, early-access feedback is consistently strong. One customer reported that the model “is the state of the art model on CursorBench. It’s opened up a class of long-horizon problems that were out of reach for earlier models.” GitHub said it “took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks.”
Fable 5’s vision capabilities extend to tasks like navigating Pokémon FireRed from start to finish using only raw game screenshots — a demonstration Anthropic uses to show how the model handles purely visual input without supplementary context. On spreadsheets, one tester found it “beats Opus 4.8 on our everyday spreadsheet suite at every effort level — and it does it with fewer turns, finishing runs 25–30% faster.”
Finance teams saw similar gains: one early user called it “the strongest finance-first model we’ve tested, both on general finance and reasoning,” and noted it “is the first to break 90% on our core analytics benchmark of complex, long-running analytical tasks — a 10-point jump over Opus.” On legal work, a separate team reported the model “feels materially different. In blind review, our lawyers found its redlines matched or beat our current model every time.”
For agentic coding, one customer reported Fable 5 “delivers more capable engineering in fewer turns than prior models — handling the complex multi-agent workflows our employees run daily in Claude Code.” On frontier physics research, another tester called it “the strongest model we’ve tested while using a third of the reasoning tokens.” On ViBench, the model was described as “the highest-performing model we’ve tested — nearly saturating our base use cases and building apps in less time with fewer tokens.”
The reasoning profile drew particular attention. One early user wrote: “Claude Fable 5’s reasoning is a clear step beyond Opus 4.8. It works at senior research scientist grade — picking directions, allocating resources, killing its incorrect beliefs, and producing novel first-principles outputs.” Another added: “At the highest effort, Claude Fable 5 reflects on and validates its own work. For us, that’s what makes highly autonomous operations possible — the extra thinking pays for itself.” For product builders: “Claude Fable 5 understands what builders mean, not just what they type. Apps that took a hundred prompts a year ago, it now one-shots.”
Concrete Gains for Early Adopters — and Real Gaps for Everyone Else
Stripe tested Fable 5 on a codebase-wide migration of a 50-million-line Ruby codebase. According to Stripe, the same task would take a whole engineering team four days — or, depending on scope, two months of manual effort — to complete by hand; Fable 5 finished in one day. Access runs through the Claude API, priced at $10 per million input tokens and $50 per million output tokens.
On the science side, Anthropic conducted drug design tests in collaboration with Dyno Therapeutics, focused on designing adeno-associated viruses (AAVs). According to Anthropic, Mythos 5 accelerated aspects of the drug design process by a factor of 10. Across 14 protein targets that yielded strong candidates, 9 are currently being actively investigated.
In molecular biology evaluations, scientists preferred Mythos 5’s hypotheses 80% of the time over competing outputs, according to Anthropic. The company also reports that Mythos 5 trained a model 100 times smaller than a comparable one published in the journal Science. One of Mythos 5’s hypotheses was independently corroborated by a study from a lab independently working on the same problem.
Business customers using Fable 5 and Mythos 5 will have their data retained for 30 days under a new policy that also applies to future models; Anthropic has published a full post on data retention policy. The company has separately invested in detecting and countering large-scale attempts to extract Claude’s capabilities.
The limitations Anthropic acknowledges are direct: without safeguards, the company says, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage. Security researchers and advanced biology teams working at the frontier will encounter fallback responses that their Mythos 5 counterparts will not. The 5% trigger rate, while low in aggregate, hits harder for professional users whose queries cluster in exactly the restricted categories.
Two Tiers, One Model — What the Competitive Picture Looks Like
The performance gap between Fable 5 and Mythos 5 is most visible in speed benchmarks. Mythos 5 reached near-parity with GPT-5.5 on frontier physics research in 36 hours — a task GPT-5.5 required four days to complete, according to Anthropic. The company claims Mythos 5 carries the strongest cybersecurity capabilities of any publicly acknowledged model, though that claim describes a product the vast majority of professionals cannot currently access.
Access to Mythos 5 runs through Project Glasswing, a 2026 program built in collaboration with the US government to provide Mythos-class models to cyber defenders and critical software infrastructure providers. Partners in Glasswing have already helped cyber defenders secure critically important software, according to Anthropic. The approximately 150 groups that held Mythos Preview access are being upgraded to Mythos 5.
For the broader market, Fable 5 is priced identically to Mythos 5 — $10 per million input tokens, $50 per million output tokens — making it a premium offering regardless of which tier a customer occupies. Anthropic says it will work to reduce safeguard false positives as more capable models arrive in the coming months, but has not committed to a specific timeline.
What Remains Unanswered
The central question Anthropic has not publicly resolved is how quickly the safeguard gaps will close. The 5% fallback rate may be acceptable for general knowledge workers, but professional security researchers and biologists working on sensitive projects face a persistent capability gap relative to Glasswing participants. Anthropic has signaled intent to expand access but has not defined the criteria for doing so.
The Glasswing model itself raises a scaling question. Roughly 150 organizations currently hold Mythos 5 access through the government partnership; independent researchers and smaller security firms may find themselves permanently in the Fable tier regardless of legitimate need. How Anthropic adjudicates expanded access — and whether that process will be transparent — remains to be seen.
Anthropic’s own acknowledgment that classifiers can produce false positives and that no bug bounty can guarantee future jailbreak immunity keeps the security calculus open. The system card and risk report offer the most complete public accounting of where the model’s boundaries currently sit — and where, by the company’s own admission, they do not.
FAQ – Frequently Asked Questions
How will Anthropic monitor and update the safeguards in Claude Fable 5?
Anthropic plans to continuously monitor Claude Fable 5’s performance through user feedback and internal testing, updating the safeguards as needed to prevent potential misuses. The company will also engage with external experts to identify potential vulnerabilities. Updates are expected to be rolled out quarterly.
Can enterprise customers customize the safeguards in Claude Fable 5 for their specific use cases?
Yes, enterprise customers can work with Anthropic to tailor the safeguards to their specific needs. This may involve adjusting the sensitivity of the classifiers or adding custom rules to handle unique use cases. Anthropic provides dedicated support for enterprise customers to help them optimize the model’s performance.
What kind of support does Anthropic offer for integrating Claude Fable 5 into existing workflows?
Anthropic provides comprehensive support for integrating Claude Fable 5, including documentation, API guides, and dedicated customer support teams. The company also offers training and onboarding programs to help customers get the most out of the model. Additionally, Anthropic has a community forum where customers can share best practices and learn from other users.
Last Updated on June 9, 2026 7:11 pm by Laszlo Szabo / NowadAIs | Published on June 9, 2026 by Laszlo Szabo / NowadAIs
