Anthropic has released Claude Fable 5, the first publicly available model in the Mythos class, and built in a safeguard system that automatically switches to a safer model across three high-risk areas.
The Mythos class denotes Anthropic's most capable model tier, running on the same base model as the newly unlocked Claude Mythos 5. Fable, by contrast, forms the new mid-tier: identical core capabilities, however with restricted access to safety-critical functions. Anthropic had already announced Claude Mythos Preview in April 2026, citing at the time excessive misuse risks for a public release. Initially, access remained limited to more than 40 elite technology companies through the Project Glasswing program. Unauthorized access to Mythos systems via a third-party environment additionally intensified the security debate. Fable 5 now replaces the previous Claude Opus 4.8 as the strongest publicly accessible model. At the same time, the provider unlocked Claude Mythos 5 with an identical base model, but with safeguards lifted in certain areas, for a restricted group of users.
Subscribe to our newsletter
The best articles of the week, directly delivered into your mailbox.
Claude Fable 5: pricing structure and availability from day one
At 10 USD per million input tokens and 50 USD per million output tokens, Claude Fable 5 costs exactly twice as much as the previous Opus 4.8, which is priced at 5 USD and 25 USD respectively. Compared with the Mythos Preview, the price drops sharply: the new mid-tier costs less than half. As a result, Mythos-class performance moves for the first time into a range that becomes calculable for a broader group of developers. Moreover, Anthropic deliberately lowers the cost of access to its strongest public model this way.
Through 22 June 2026, the model is additionally included free of charge in the Pro, Max, Team and Enterprise plans. Via the API, Fable 5 is initially fully available right away, so developers can access the new tier without delay. Furthermore, the model has been generally available in GitHub Copilot since 9 June 2026. Consequently, this broad distribution across multiple channels brings Mythos-class performance quickly into production development environments.
Safeguard architecture: automatic fallback across three risk areas
The core of Fable 5 is a mechanism that automatically routes requests in three defined areas to Claude Opus 4.8. The areas affected are cybersecurity with exploitation and agent-based hacking, biology and chemistry with dual use, and distillation, where users might attempt to extract model capabilities. According to Anthropic, this fallback occurs in fewer than 5% of sessions. In more than 1,000 hours of external red-teaming, testers additionally found no universal jailbreaks.
Alongside this, Anthropic introduces a new data protection rule: Mythos-class traffic is subject to a 30-day retention period, and the data does not feed into model training. Furthermore, the company positions itself as the first provider with a formal framework for the controlled release of models it classifies as too dangerous for the unrestricted public. Thus the question shifts from a pure release decision toward a tiered access system.
Claude Mythos 5 serves as the counterpart, with safeguards lifted in individual areas. Access goes to selected partners through Project Glasswing, among them AWS, Microsoft, Apple and CrowdStrike. The model is likewise open to selected bio-researchers, for whom the restrictions in the biological area no longer apply.
Benchmark comparison: Claude Fable 5 against Opus 4.8 and competitors
On the relevant software engineering benchmarks, Fable 5 sets itself clearly apart from its predecessor. According to SWE-bench Verified, the model reaches 95.0% against 88.6% for Opus 4.8, a lead of 6.4 percentage points. On the more demanding SWE-bench Pro, the gap is even more pronounced at 80.3% against 69.2%. The competition trails far behind here: GPT-5.5 reaches 58.6%, Gemini 3.1 Pro 54.2%.
The leap also shows in specialized tests. On FrontierCode Diamond, Fable 5 scores 29.3% against 13.4%, more than double its predecessor. On the Hex Core analytics benchmark, it is furthermore the first model ever to reach a hit rate of 90%. In practice, Stripe migrated a Ruby codebase of more than 50 million lines of code in a single day instead of the estimated two months. Finally, Fable 5 benefits three times more from persistent memory than Opus 4.8.
Why Anthropic considers the safeguards essential becomes clear from the ExploitBench of the unconstrained Mythos 5. There the model reaches 78.0% against 40.0% for Opus 4.8 and only 34.0% for GPT-5.5. Therefore, this gap in offensive cyber capabilities explains in hindsight why the provider anchored the automatic fallback in the public version.
Anthropic IPO near the 1-trillion mark as context for investors
The product release falls into a decisive phase: Anthropic confidentially filed a draft S-1 form with the SEC on 1 June 2026. The latest funding round of 65 billion USD valued the company at 965 billion USD, just below the threshold of one trillion. Goldman Sachs, JPMorgan Chase and Morgan Stanley are additionally seen as bookrunners for the listing.
The valuation rests on a steep growth curve. The revenue run rate stood at around 47 billion USD in May 2026 and is expected to cross the 50 billion USD mark by the end of June. For comparison: in July 2025, the run rate was still 4 billion USD. Thus the running revenue has grown more than tenfold in less than a year.
Notably, the timing relative to its competitor stands out: OpenAI likewise filed confidentially for a listing on 8 June 2026. This parallel IPO wave by the two largest AI providers is seen as a possible signal of sector maturity. For investors, Anthropic's valuation consequently acts as a proxy for the valuation level of the entire AI market, while Fable 5 serves as a product catalyst within the IPO window.
