On June 9, Anthropic opened public access to Claude Fable 5 – the first Mythos-level model deemed safe for everyone. Before this, the Mythos technology was kept under lock and key in a closed program since spring due to overly strong cyber capabilities. Now it is given to developers and subscribers, but with protective restrictions. In parallel, Claude Mythos 5 was released for a narrow circle of partners. Let's break it down in order: what this model is, what it shows in practice, how much it costs, and what the catch is for business.
What is Claude Fable 5
Claude Fable 5 is the public flagship of the new Claude 5 generation. According to Anthropic, its capabilities surpass any model the company has released to the public. It shows top results on almost all tested benchmarks: development, analytics, vision tasks, and scientific research. The pattern is simple: the longer and more complex the task, the more noticeable the gap between Fable 5 and other models.
The main thing to understand from the very beginning: Fable 5 and Mythos 5 are the same base model. The only difference is in the safeguards. For the public Fable, they are enabled; for the partner Mythos, they are removed in the field of cybersecurity. Even the names refer to this: Fable from the Latin fabula, "that which is told", is related to the Greek mythos. It is the restrictions that distinguish the models, which is why the names are different.
Where the model came from: a brief timeline
To understand the scale of the release, context is needed. The existence of Mythos became known in March 2026 due to a leak of Anthropic blog drafts. The company then confirmed the development and called it a qualitative leap. In April, Mythos Preview was launched in closed mode through the Project Glasswing program – only for organizations involved in cybersecurity and critical infrastructure protection. The model was not released to the public: its ability to find vulnerabilities and build exploit chains was deemed too dangerous.
A few days before the official release, an identifier appeared in the price list of a third-party aggregator claude-oceanus-v1-p. It was an early preview checkpoint of the same model distributed to testers. We have compiled the full timeline of that leak in a separate article.
What Fable 5 shows in practice
Anthropic provided specific case studies from early clients, and they speak louder than dry percentages.
Stripe reported that Fable 5 compressed months of engineering work into days. In a codebase of 50 million lines of Ruby, the model migrated the entire base in one day – manually, it would have taken the team more than two months. On the Cognition FrontierCode benchmark, which tests high-quality maintainable code, Fable 5 showed the best result among frontier models even at a medium level of effort.
In vision tasks, the model sets a new bar. It reconstructs the source code of a web application from screenshots alone and completes Pokémon FireRed on a minimal vision harness – where previous Claude models stalled even with auxiliary tools.
Memory and long context are highlighted separately. In the game Slay the Spire, access to file memory improved Fable's performance three times more than Opus 4.8, and it reached the final act three times more often.
How Fable differs from the closed Mythos: safeguards
The main difference of the public version is the classifier system. These are separate AIs that catch potentially dangerous requests. When a Fable classifier triggers on a topic from three categories – cybersecurity, biology and chemistry, or an attempted distillation – the response is automatically routed to the less powerful Claude Opus 4.8 model, and the user is notified.
According to Anthropic, such a fallback happens rarely: in over 95% of sessions, Fable works without switching. For these sessions, its performance is practically equal to the full Mythos 5. The safeguards are intentionally set strictly, so sometimes harmless requests fall under them. The company acknowledges this and promises to reduce false positives after launch.
Why so strict. Mythos-level models can find and exploit vulnerabilities, which makes cyberattacks cheaper and easier. For biology and chemistry, the logic is the same: the same requests are useful to a researcher and dangerous in the hands of an attacker. Therefore, for now, Fable falls back to Opus 4.8 on most requests from these areas.
Anthropic conducted an external bug bounty – more than 1000 hours of testing, and during this time no one found a universal way to bypass the protection.
How much Claude Fable 5 costs
The price of both models via API is the same:
Input: 10 dollars per 1 million tokens
Output: 50 dollars per 1 million tokens
This is twice as expensive as Opus 4.8, which is 5 and 25 dollars. But at the same time, it is less than half the price of the previous Mythos Preview, which cost 25 and 125 dollars. That is, Anthropic not only opened the Mythos class to the public, but also made it significantly cheaper than it cost in the closed version.
For subscribers, the procedure is as follows. From June 9 to 22, Fable 5 is included in the Pro, Max, Team, and seat-based Enterprise plans at no extra charge. From June 23, the model will be removed from these plans, and usage credits will be needed thereafter. When there is enough capacity, Anthropic plans to return Fable 5 to standard plans. On API and consumption-based Enterprise, the model is fully available from day one.
Important for developers
If you work with the Claude API, there is a technical detail that is missing in most summaries. The API surface of Fable 5 matches Opus 4.7 and 4.8, but with one breaking change: the explicit parameter thinking: {type: "disabled"} now returns a 400 error. If you do not need the extended reasoning mode – simply do not pass this parameter, rather than disabling it explicitly. The model identifier in the API is claude-fable-5.
Another point that directly concerns business. For all Mythos-level models, mandatory 30-day data retention is introduced – on all platforms, including third-party ones. This data is not used for training and is deleted after 30 days, but Anthropic needs it to protect against complex attacks and reduce false positives of classifiers. Those with strict data processing requirements should factor this into their assessment.
What this changes in practice
In short – a Mythos-level model is available to regular teams for the first time, and not just to special partners.
Multi-step agent scenarios get a more reliable foundation. Opus 4.8 already handles long task chains quite well, but Fable 5 retains context and planning significantly better over a distance of millions of tokens. Working with large codebases is accelerated, which is confirmed by the Stripe case study.
At the same time, expect that in sensitive areas the model will refuse more strictly and switch to Opus 4.8. For legitimate tasks this is rarely a problem, but if your work is in cybersecurity or biology – some requests will fall back.
Frequently asked questions
What is Claude Fable 5?
This is the first public Anthropic model of the Mythos tier and the first in the Claude 5 lineup. Released on June 9, 2026, it is available via the Claude API and paid subscriptions.
How much does Claude Fable 5 cost?
10 dollars per 1 million input tokens and 50 per 1 million output tokens. This is twice as expensive as Opus 4.8 and less than half the price of the closed Mythos Preview.
How does Fable 5 differ from Mythos 5?
It is the same model. Fable has public safeguards enabled, while Mythos has them removed in the field of cybersecurity – Mythos is available only to Project Glasswing partners.
What are the limitations of Fable 5?
In cybersecurity, biology, chemistry, and during distillation attempts, the model does not answer itself, but passes the request to the Opus 4.8 model. This triggers on average in less than 5% of sessions.
What is better – Fable 5 or Opus 4.8?
Fable 5 shows top results on almost all benchmarks and in the Stripe case study migrated a 50 million line codebase in a day. But it costs twice as much, so for simple tasks Opus 4.8 remains a rational choice.
Summary
Claude Fable 5 – this is the moment awaited since spring: Mythos technology has moved beyond closed programs, and is even cheaper than it cost in the closed version. With safeguards in cybersecurity, biology, and chemistry, and at twice the price of Opus, but in the hands of regular developers and subscribers. This is the first model in the Claude 5 lineup, and it significantly raises the bar. Next in line is OpenAI's response with GPT-5.6.
