Anthropic released Claude Fable 5 on Tuesday, the first model from its Mythos class made available to the general public, the same family of models the company had previously argued was too dangerous for general release. The version now in subscribers’ hands routes any query touching cybersecurity, biology, chemistry, or model distillation through a classifier, and quietly hands it to Claude Opus 4.8 if the classifier fires.
The ‘Mythos-Class’ Pitch
Anthropic calls Fable 5 a ‘Mythos-class’ model made ‘safe for general use,’ and says its capabilities exceed those of any model it has ever made generally available, in its joint announcement of Fable 5 and Mythos 5. Fable 5 is ‘state-of-the-art on nearly all tested benchmarks of AI capability,’ with exceptional performance in software engineering, knowledge work, vision, and scientific research. The longer and more complex the task, the wider the gap between Fable 5 and Anthropic’s other models, the company says.
Both Fable 5 and the parallel Mythos 5 release are priced at $10 per million input tokens and $50 per million output tokens, less than half the price of Claude Mythos Preview, per Anthropic. The Claude app and Pro, Max, Team, and seat-based Enterprise plans include Fable 5 at no extra cost through June 22, with the Claude API and consumption-based Enterprise plans getting it from launch day. The model also ships on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. After June 23, subscriber access shifts to usage credits billed at API rates, with the model returning to standard plans when capacity allows.
| Aspect | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Initial access | Public, Pro/Max/Team/Enterprise through June 22, then usage credits | Project Glasswing partners only |
| Safeguards | Three classifier domains, fallback to Claude Opus 4.8 | Safeguards lifted in some areas |
| Pricing (API) | $10 per million input, $50 per million output | $10 per million input, $50 per million output |
From Glasswing Lockdown to a Pro Subscription
Two months ago, the Mythos class was, by Anthropic’s own framing, too risky to ship at all. The company launched Project Glasswing on April 7, 2026, an initiative that brought together AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to use a frontier model called Claude Mythos Preview for defensive cybersecurity work. Anthropic committed up to $100M in usage credits for the project and $4M in direct donations to open-source security organizations.
Mythos Preview, the Glasswing version, found thousands of high-severity vulnerabilities, including some in every major operating system and every major web browser, Anthropic says. The model’s standout catches: a 27-year-old remote crash in OpenBSD, a 16-year-old bug in FFmpeg that automated testing had hit five million times without catching, and a chain of Linux kernel exploits that escalated from ordinary user access to full machine control, per the Glasswing technical writeup. On CyberGym, a cybersecurity vulnerability reproduction benchmark, Mythos Preview hit 83.1% against Claude Opus 4.6’s 66.6%, a 16.5-point gap in the same tier of model. The pattern of AI finding decade-old bugs in critical open-source projects has only accelerated since, with a smaller team using cheaper models reproducing the same depth at a tenth of the cost, and Anthropic has since expanded Project Glasswing to 150 new organizations in more than fifteen countries, again focusing on groups that maintain critical infrastructure.
The same underlying model that was held back from public release two months ago is on a Pro subscription as of Tuesday. It ships wrapped in classifier safeguards.
How the Safeguards Work
Fable 5 ships with three classifier domains active, covering cybersecurity, biology and chemistry, and model distillation, the last one aimed at stopping competitors from extracting Fable 5’s capabilities into their own systems. When a query trips a classifier, the response is generated by Claude Opus 4.8, Anthropic’s next-most-capable model, instead. Anthropic says the classifiers are tuned conservatively, and trigger in less than 5% of sessions on average, but they will sometimes catch harmless requests.
Anthropic ran an external bug bounty for more than 1,000 hours that produced no universal jailbreak, and external red-teaming organizations also failed to find one, per TechCrunch’s coverage of the launch. A mandatory 30-day retention on all Fable 5 and Mythos 5 traffic, even for enterprises with prior zero-retention agreements, is in place to defend against novel jailbreaks and to reduce false positives in the classifiers, per the same report. Anthropic says the data is not used for training. The policy applies to every Mythos-class prompt, including ones that never trip a classifier.
- Launch date: June 9, 2026
- Class: Mythos-class
- Fallback model: Claude Opus 4.8
- Classifier trigger rate: less than 5% of sessions
- Subscription access: free through June 22, then usage credits
What the Benchmarks Show
Anthropic’s claim is that Fable 5 leads on nearly every benchmark the company tested, and that the lead widens as task complexity grows. The numbers Anthropic chose to publish lean on third-party workload proofs rather than academic test suites. On the agentic coding side, Stripe reported that Fable 5 performed a codebase-wide migration of a 50-million-line Ruby codebase in a single day, work that would otherwise have taken a full engineering team over two months by hand.
Fable 5 scored 80.3% on SWE-Bench Pro, against GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%. On Hebbia’s Finance Benchmark for senior-level reasoning, the model took the top score, with Anthropic pointing to substantial gains in document-based reasoning, chart and table interpretation, and problem solving. Analytics platform Hex said Fable 5 was the first model to break 90% on its core analytics benchmark of complex, long-running tasks, a 10-point jump over Opus.
On Cognition’s FrontierCode evaluation, which tests whether models can pass difficult coding tasks while meeting the standards of high-quality production codebases, Fable 5 scored highest among frontier models even at medium effort. For vision, Fable 5 beat Pokémon FireRed using only a minimal, vision-only harness, with no maps, no navigation aids, and no extra game-state information, after earlier Claude models had failed to finish the game even with complex helper harnesses. With file-based memory enabled in Slay the Spire, Fable 5’s performance improved three times more than Opus 4.8’s, and it reached the game’s final act three times more often.
On the science side, Mythos 5 (the restricted, safeguard-lifted sibling) accelerated the drug design process by around ten times, per Anthropic’s internal protein design experts, and nine of 14 protein targets in one study yielded strong candidates for drug design. In blinded head-to-head comparisons, Anthropic scientists preferred Mythos 5’s molecular biology hypotheses around 80% of the time against Opus-class models, and one Mythos hypothesis, a novel mechanism for an E. coli protein, was corroborated in a separate study from a lab independently working on the same problem.
The 30-Day Retention Mandate
Beyond the classifier architecture, the launch also brings a new mandatory data-retention policy. With Fable 5 and Mythos 5 live, Anthropic is requiring a 30-day retention on all traffic, even for enterprises that previously had zero-retention agreements, per TechCrunch’s coverage of the launch. Anthropic says the retained data will not be used for training.
It will be used to defend against novel jailbreak attempts and to identify and reduce false positives in the safety classifiers. The policy lands on every Mythos-class prompt, including ones that never trip a classifier.
For Fable 5 users, the practical consequence is that ‘which model answered’ is now a probabilistic question, since the user cannot always tell when a query has been routed to Opus 4.8. The policy applies across all subscription tiers and across all platforms where Fable 5 ships, including the Claude app, the API, AWS, Bedrock, Vertex AI, and Microsoft Foundry. For a model whose whole selling point is the same underlying weights as the restricted version, the mandatory 30-day retention of every prompt is a privacy compromise even with the no-training promise attached. The data retention will be tested as Mythos-class models roll out to more enterprise customers in the months ahead.
Mythos 5 for Defenders, Fable 5 for the Rest
For a small group of cyberdefenders and infrastructure providers, Anthropic launched Claude Mythos 5 alongside Fable 5. The two share underlying weights, per Anthropic’s launch post. The difference is that Mythos 5 ships with some of the safeguards lifted, and is initially available only through Project Glasswing, in collaboration with the US government, as an upgrade to Claude Mythos Preview, which Anthropic says has the strongest cybersecurity capabilities of any model in the world.
Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage.
That is Anthropic’s own framing, in the launch post for the model. Anthropic says it intends to expand access to Mythos 5 ‘soon’ through a broader trusted access program. The biology and chemistry safeguards may also be relaxed for vetted biology researchers through a future program, per the company. The deployment pattern, a single model with three trust tiers, a fallback architecture, and a partner-managed safety net, is new to frontier lab releases.
Frequently Asked Questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s first publicly available Mythos-class model, released on June 9, 2026. The Mythos class is Anthropic’s most capable model tier, previously held back from general release because of its cybersecurity capabilities.
Is Claude Mythos 5 the same model as Fable 5?
Yes, the two models share underlying weights, per Anthropic’s joint launch post. The difference is that Fable 5 ships with classifier safeguards active, while Mythos 5 has those safeguards lifted in some areas. Mythos 5 is initially available only to cyber defenders and infrastructure providers through Project Glasswing.
How much does Fable 5 cost?
Fable 5 is priced at $10 per million input tokens and $50 per million output tokens on the Claude API and consumption-based Enterprise plans, less than half the price of Claude Mythos Preview. For Pro, Max, Team, and seat-based Enterprise subscribers, Fable 5 is included at no extra cost through June 22, 2026.
When does free Fable 5 access end for subscribers?
Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026. On June 23, Anthropic will remove Fable 5 from those plans, and continued use after that date requires usage credits billed at API rates. The model will return to standard plans when capacity allows, per Anthropic.
What are the new safeguards on Fable 5?
Fable 5 runs three classifier domains, covering cybersecurity, biology and chemistry, and model distillation. When a query trips a classifier, the response is generated by Claude Opus 4.8 instead, and Anthropic says the classifiers fire in less than 5% of sessions on average.
What is Project Glasswing?
Project Glasswing is an Anthropic-led cybersecurity initiative announced on April 7, 2026. It brought together AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Its purpose is to use Mythos-class models to find and fix vulnerabilities in critical software, the same capability that, uncontained, made Anthropic hold the model back from public release.
