Anthropic releases Claude Fable 5 broadly, rerouting risky requests to Opus 4.8

Anthropic just opened the floodgates on a new Claude line-up. Starting today, Fable 5 is broadly available, and it is positioned as the most powerful Claude model that is generally available so far. Anthropic says Fable 5 exceeds every Claude model it previously made generally available, with stronger performance across software engineering, knowledge work, vision, scientific research, and long-running tasks. Meanwhile, the prior “Mythos-class” capability did not previously go wide. It was available only through a restricted cybersecurity program, Project Glasswing, which Anthropic announced two months ago.

The real twist is how Anthropic is trying to make that frontier capability safer without throttling it. For certain high-risk areas, Anthropic says requests are automatically routed to Claude Opus 4.8 instead, and users get notified when that routing happens. Those high-risk areas include cybersecurity, biology, and chemistry, plus attempts to distill the model’s capabilities into other systems. Anthropic adds that more than 95% of Fable sessions run entirely on Fable’s own responses, with no fallback. Translation: most “normal” work runs on the new general-purpose model, but the risky edge cases do not.

This is not just a model release. It is a deployment philosophy. Anthropic is effectively selling “power by default, control by design.” That matters because dual-use capabilities are where regulators, enterprises, and customers all start pulling in different directions. The source frames Anthropic’s approach as bringing frontier models with dangerous dual-use capabilities into the market, not by releasing everything to everyone, and not by simply refusing risky questions. Instead, the company routes some requests to a less capable model while keeping the stronger model available for the majority of everyday work.

On pricing and access, Anthropic is also making a bet that enterprises and developers will adapt quickly. It priced both Fable 5 and Mythos 5 at $10 per million input tokens and $50 per million output tokens. Anthropic claims that is less than half the price of Claude Mythos Preview, but it still makes it the most expensive major AI models available globally, according to a pricing snapshot included in Anthropic’s materials. For developers, Fable 5 is available through the Claude API as claude-fable-5, with full availability today on the Claude API and on consumption-based Enterprise plans. For subscription users, the rollout is time-boxed: Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from today through June 22. On June 23, Anthropic plans to remove Fable 5 from those plans, after which usage credits will be required. The company says it aims to restore Fable 5 as a standard part of subscription plans as quickly as possible.

Now the distinction between Fable 5 and Mythos 5: Anthropic is not presenting them as a simple small-versus-large split. The source says they appear to share the same underlying base capability level. The difference is access control and embedded guardrails. Mythos 5 lifts some restrictions for trusted users working in approved domains. In practical terms, Anthropic says Mythos 5 is more powerful for sensitive cyber and biology work because it can answer in areas where Fable 5 falls back. For most ordinary enterprise and developer tasks, Anthropic says Fable 5 performs effectively the same as Mythos 5. So Mythos 5 is not “more magic” across the board. It is more permissioning in the specific places where Anthropic’s routing layer would otherwise kick in.

Where Fable 5 is most immediately compelling to enterprise buyers is software engineering and agentic coding. Anthropic says Fable 5 can work unattended for longer and with more independence than previous Claude models, which is the capability enterprises need if they want AI agents to do more than autocomplete code or answer developer questions. On SWE-bench Pro, which measures a model’s ability to complete difficult software engineering tasks, Anthropic says Fable 5 and Mythos 5 reach 80.3%, compared with OpenAI’s GPT-5.5 at 58.6%. On Cognition’s FrontierCode Diamond benchmark, testing high-quality, maintainable agentic coding, Anthropic reports 29.3% for the models versus 13.4% for Claude Opus 4.8 and 5.7% for GPT-5.5, based on the benchmark table included in Anthropic’s materials.

Anthropic also cites something practical: Fable 5 scores highest among frontier models on FrontierCode even at medium reasoning effort, suggesting stronger coding results without always needing maximum compute. That matters for cost predictability and for teams trying to operationalize agents. The customer examples in the source lean hard into “autonomy with throughput.” Stripe tested Fable 5 in a 50-million-line Ruby codebase and found it completed a codebase-wide migration in one day that otherwise would have taken a team more than two months by hand. Cursor says “Fable 5 is the state of the art model on CursorBench,” and that it opened up a class of long-horizon problems out of reach for earlier models. Replit says Fable 5 is the highest-performing model it has tested on ViBench, its end-to-end “vibe-coding” benchmark, and that it builds apps in less time with fewer tokens. Figma describes it as a “clear step forward on agentic coding and prototyping,” and other early users cited in the source include Base44, Genspark, and Rakuten, each emphasizing deeper autonomy, full-app one-shotting, strong eval wins, and reflection or validation at the highest effort.

Strategically, this launch is a roadmap for how frontier AI could enter enterprises without detonating risk controls. Routing is becoming a product feature, not just safety engineering. For boards and leaders, the stake is whether AI “agents” become a genuine operating capability or stay a demo. Anthropic’s message is clear: if you can keep most day-to-day work on the strongest generally available model, while shifting the most dangerous slices to a safer fallback, you get both adoption speed and risk containment. That is the chess move peers will be watching as model access, agent reliability, and enterprise governance all converge.