Claude Fable 5: Anthropic has released a public version of Mythos – the most powerful model in the open access
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
On June 9, 2026, Anthropic released the *Claude Fable 5, the first publicly available model from the new Mythos class. This is a direct modification of the system, which the company in April categorically refused to release to a wide audience, citing cybersecurity risks. Within two months, Anthropic had developed a barrier architecture that it felt was sufficient to open up access for paying subscribers and enterprise customers.
What is Mythos and why the model was kept locked
In April 2026, Anthropic announced Claude Mythos Preview as part of Project Glasswing. The company did not plan to make the model public, limiting itself to a narrow circle of partners who were engaged in cybersecurity. The reason is unprecedented opportunities in the search and exploitation of vulnerabilities.
Mythos has demonstrated an exceptional ability to autonomously locate zero-day exploit chains in large operating systems and browsers. Project Glasswing members — including AWS, Microsoft, Apple, and CrowdStrike — have discovered thousands of serious vulnerabilities during their work with the model.
The public release of such a system without restrictions created an obvious problem: the same tools that help defend can be used for attacks. This is exactly what Anthropic has been working on for the past two months.
Fable 5 vs Mythos: What's the difference
Mythos-class models are a separate level of capabilities above the Opus class. The first was Claude Mythos Preview, released in April through Project Glasswing. Today, it is followed by Claude Fable 5 and Claude Mythos 5. Fable comes from the Latin fabula - "that which is told" related to Greek mythos. It is the protective barriers that distinguish the two models - Fable and Mythos - and that is why they have been given different names.
Simply put, it’s one architecture, but with different levels of constraints.
Claude Fable 5 is a public version with security mechanisms, locked themes and automatic switching to Opus 4.8 in dangerous areas. Available to paid claude.ai subscribers and via API.
Claude Mythos 5 is an updated version for Project Glasswing members. Anthropic is simultaneously rolling out a new version of Mythos – Mythos 5 – for organizations that have already gained access to the advanced model. This version is much less limited in the field of cybersecurity.
How protective mechanisms work
The main technical feature of Fable 5 is its hybrid response architecture. Without security barriers, Fable 5’s capabilities in areas like cybersecurity could be exploited to the detriment. Therefore, the model launches with mechanisms in which queries on certain topics will be redirected to the next most powerful model – Claude Opus 4.8.
In high-risk areas—cybersecurity, biology, chemistry, distillation—the model blocks answers and switches to Claude Opus 4.8.
In practice, it looks like this: the user asks about the creation of the toxin - Fable 5 intercepts the request and responds through Opus 4.8. The remaining 95% of the sessions go through Fable 5.
For a fast and secure release, Anthropic has configured these protections conservatively - they sometimes intercept innocuous requests, although on average they work in less than 5% of sessions. The company plans to reduce the percentage of false positives as more accurate classifiers are developed.
Hacking testing
The company conducted a large-scale internal red-teaming of its classifiers, after which it launched an external bug bounty program lasting more than 1,000 hours, which revealed no universal jailbreaks. Independent external red-teaming also found no critical bypasses.
This is a serious result: previous Anthropic models were hacked. An academic paper by STACK, published in June 2026, described a method of bypassing Claude Opus 4 classifiers with a success rate of 79% on bioweapon requests. Fable 5 did not respond to similar testing.
Opportunities: What the model can do
According to the company, “Fable’s capabilities surpass any model they’ve ever released.” It is at the forefront of almost all benchmarks tested, demonstrating exceptional performance in software development, knowledge, computer vision, research and many other areas.”.
Hex said in a statement that the Fable was the first model to score 90% on their baseline analytics benchmark for complex, long-term analytics tasks.
Anthropic’s emphasis is on long-term multi-step tasks, where the increase from the previous generation is most noticeable. This directly concerns vibcoding: agent scripts with Claude Code, long refactorings, complex research tasks.
Price: more expensive Opus, cheaper than expectations
Fable costs about twice as much as the current flagship models Claude Opus. The original Mythos price, announced in April, was five times the Opus price. The public version is cheaper, but still double what was available before.
If you focus on current API prices – Claude Opus 4.7 costs $5/$25 per million tokens (input/output) – then Fable 5, according to available data, will cost about $10/$50. The official prices of Anthropic at the time of publication are updated.
Mandatory storage of traffic: a new precedent
With the launch of Fable 5 and Mythos 5, Anthropic introduces mandatory 30-day storage of traffic, even for those enterprise customers who previously had zero-storage agreements. The company said it would not use the data to train models - only to "protect against complex and new attacks, including new jailbreaks" and "identify and reduce false positives.".
This is a major change for enterprise customers with strict privacy requirements. The policy could set a precedent in which access to more powerful models will cost query data - even if that data is not used for training.
Project Glasswing and Mythos 5
Simultaneously with the public launch of Fable 5, Anthropic announced that Project Glasswing partners are getting access to the updated Mythos 5.
Last week, Anthropic expanded access to Mythos to hundreds of organizations in 15 countries, focusing on organizations managing critical infrastructure. The company plans to gradually expand the privileged access program through a structured verification system.
Context: IPO and power race
The launch of Fable comes amid preparations for Anthropic to enter public markets - simultaneously with OpenAI and Elon Musk's SpaceX. It also follows the company's call for the world's largest AI labs to create a coordinated mechanism to inhibit the development of frontline systems. Anthropic warned that systems are evolving so rapidly that they could soon achieve recursive self-improvement - the ability to autonomously improve themselves without human intervention.
Here is the central contradiction of Anthropic’s position: a company that publicly warns about the existential risks of AI is simultaneously releasing the most powerful publicly available model in history. “It’s about what we call the race to the top – the ability to deliver this technology in a valuable way while creating the right protective barriers so that it does asymmetrically more good than harm,” said Dianne Penn, director of product management for research at Anthropic.
What it means for developers
For vibcoding and agent tasks Fable 5 is a direct improvement. More powerful than Claude Code, better multi-step tasks, higher quality of complex refactorings. If you’re using the Opus layer now, Fable 5 will be noticeably better at twice the price.
** For enterprise customers with zero-storage agreements**, the policy change needs to be reviewed. 30-day storage of traffic when working with Fable 5 and Mythos 5 is now a must.
For cyber security and biology applications, Fable 5 is limited and automatically switches to Opus 4.8. If you want full Mythos class capabilities in these areas, you will need to apply for Project Glasswing.
API-line model at the time of publication is not officially announced. Stay tuned for updates on docs.anthropic.com.
