Claude Fable 5: Why Anthropic Released a Model Long Considered Too Dangerous

Claude Fable 5: why Anthropic released a model considered too dangerous for months

In April 2026, Anthropic introduced Claude Mythos and immediately restricted access to it: only for selected organizations involved in critical infrastructure protection. On June 9, the same company announced Claude Fable 5 — a public version of the model based on the same technological foundation. What has changed in two months, and why now?

What is Claude Fable 5

Claude Fable 5 is the first publicly available model of Anthropic's new Mythos class. The name is not accidental: fable from Latin is fabula, "that which is told" — the same root as the Greek mythos. This is not a marketing game, but a signal: Fable 5 and Mythos 5 are built on the exact same model weights. The difference is not in the architecture, but in the safety layer.

Anthropic positions the Mythos class as a level above Opus. The first model of this class, Claude Mythos Preview, was released in April 2026 through the closed Project Glasswing — and remained unavailable to the general public until June 9.

Why was the Mythos model hidden from the public

The main reason is cybersecurity. Back in April, Anthropic described Mythos's capabilities as being able to "extremely effectively detect and exploit vulnerabilities in software." It was this characteristic that impressed both Wall Street and US government structures — and it was precisely because of this that the company did not plan to release the model to the public.

The access scheme unfolded in stages:

• April 2026 — Mythos Preview available only to a select few partners within Project Glasswing, exclusively for critical infrastructure protection.
• End of May 2026 — the program expanded to several hundred organizations in 15 countries, but again, only for defensive tasks.
• June 9, 2026 — release of Claude Fable 5 to the general public.

Concurrently, the company publicly warned: frontier models are becoming so powerful that in the coming months they may reach the level of recursive self-improvement (RSI) — that is, autonomous improvement of their own code without human intervention. At first glance, this contradicts the decision to release Fable 5 now. But Anthropic's logic is different: not to wait until the model becomes even more powerful, but to release it with a verified defense layer — while control is still possible.

Fable 5 vs Mythos 5: what's the difference

The key point is: it's the same model. The difference is exclusively in the configuration of the safety classifiers. Fable 5 has an active defense layer; Mythos 5 is the same architecture but with partially removed restrictions for verified partners of Project Glasswing.

The table below shows ExploitBench: an industry benchmark that measures the model's ability to find and exploit real-world software vulnerabilities. This specific metric was the main reason for the limited release of Mythos.

Anthropic confirmed: on benchmarks not related to blocked domains, the difference in results between Fable 5 and Mythos 5 is 1–3 percentage points — within statistical noise.

Benchmarks: The Numbers and What They Mean

Below are key benchmarks from Anthropic's official announcement and independent tests as of June 9–10, 2026. The ★ symbol indicates that the result belongs to Mythos 5; Fable 5 drops to the level of Opus 4.8 in these categories due to fallback.

What to understand from the numbers:

• SWE-Bench Pro is a more reliable benchmark than SWE-Bench Verified. OpenAI has acknowledged data contamination in the Verified version; the Pro version tests real-world tasks on closed repositories.
• FrontierCode is the most representative for production code. 29.3% compared to 13.4% for Opus 4.8 means Fable 5 solves more than twice as many complex engineering tasks.
• ExploitBench is a mirror of the architectural decision. Mythos 5 scores 78%, Fable 5 scores close to zero. This is the essence of the distinction between the two models.

For a more detailed breakdown of previous generation benchmarks and how Opus 4.8 performed before the release of Fable 5, see our article Claude Opus 4.8: Benchmarks, Numbers, and What They Mean.

What Limitations Does the Public Version Have

Anthropic has implemented a classifier layer that intercepts requests in four categories and redirects them to Claude Opus 4.8:

• Cybersecurity — vulnerability searching and exploitation, offensive software development
• Biology — synthesis of dangerous substances, pathogen design
• Chemistry — synthesis of dual-use substances
• Model Distillation — attempts to replicate internal weights through outputs

Technically: when the classifier triggers, the API returns an HTTP 200 with stop_reason: "refusal" and a stop_details.category field with the value "cyber", "bio", "reasoning_extraction", or null. If no output was generated, the request is not charged.

According to Anthropic, the classifiers trigger in less than 5% of sessions on average. The company acknowledges the existence of false positives and promises to reduce them with new version releases.

Before the release, Anthropic conducted an external bug bounty: over 1,000 hours of testing did not reveal any universal jailbreaks.

For Which Tasks Does Anthropic Recommend Fable 5

For routine tasks where SWE-Bench Pro level complexity is not required, Claude Opus 4.8 remains a more rational choice: twice as cheap for the same tokens.

How to access Claude Fable 5

Practical tip: if the model hasn't appeared in your list yet, I'd recommend simply refreshing the page or logging out and back in. In my case, Claude Fable 5 appeared after reloading the interface, without any additional settings.

How much does Claude Fable 5 cost

If you're using LLMs for programming, it's worth looking not only at the price per token but also at the conditional cost of getting the correct result. In such evaluations, Fable 5 might appear more cost-effective than Opus 4.8 due to a higher success rate, which reduces the number of retries and overall costs for solving a task.

What the Fable 5 release means for the AI market

In my opinion, the main issue here is not technical. If just this spring a Mythos-class model was considered too risky for broad access, what exactly has changed in a few months? Anthropic's answer is not that the model has become less powerful, but that the company has begun to trust its own safety layer. According to Anthropic, an external bug bounty program could not bypass these mechanisms even after over 1,000 hours of testing and attack attempts. In fact, we are seeing the first major precedent where a frontier model is released to the public with a clearly documented system of restrictions and automatic fallback to a less risky model for specific categories of requests.

After analyzing the Fable 5 release, I've noticed several trends that could be important for companies implementing AI solutions or evaluating the strategies of major AI vendors.

• Safety mechanisms instead of bans are becoming the new standard. OpenAI, Anthropic, and Google are gradually moving away from the "release or not release" approach. Instead, they are building multi-layered control systems that allow access to powerful models while restricting specific risky scenarios. Fable 5 is currently one of the most transparent examples of this approach.
• The gap between public and private versions of models is becoming structural. Mythos 5 and Fable 5 demonstrate a new model of AI distribution. Formally, it's one technological platform, but with different access levels and different capabilities in specific domains. It's highly likely that we'll see a similar scheme in future generations of frontier models.
• Commercialization and security are no longer contradictory. For Anthropic, the Fable 5 release is not just a technological achievement but also a business signal. The company is demonstrating that it can bring the most powerful models to market without completely abandoning its safety principles. For investors, this is an important indicator of platform maturity.
• Mythos-level capabilities will gradually become the new norm. In my opinion, within the next 6-12 months, capabilities that are currently associated with the Mythos class will begin to be perceived as standard for flagship models. A similar process already occurred after the emergence of GPT-4: what initially seemed like a breakthrough became a basic market expectation a year later.

If we view the Fable 5 launch not as another model release, but as a signal to the entire industry, the main conclusion for me is this: major AI companies are no longer trying to hide their most powerful models from the general public. Instead, they are investing in control mechanisms that allow them to open access to frontier models without losing control over the most risky usage scenarios.