Anthropic's Claude Fable 5 Beats Every Rival on Coding and Research

Anthropic has launched Claude Fable 5, its most powerful model ever made publicly available. It is the first member of the Mythos class , a new capability tier that sits above Opus , to reach general release. Alongside it, Anthropic also launched Claude Mythos 5, the same underlying model but with safety restrictions lifted, reserved for a small group of vetted cyberdefenders through Project Glasswing.

Mythos-class models are a tier of Claude models that sit above the Opus class in capability. The distinction between the two new models comes down to one thing: safeguards. Fable 5 is the public version with safeguards, while Mythos 5 is the same model with some protections removed, reserved for cybersecurity professionals via Project Glasswing.

Two months in the making

Anthropic first introduced Claude Mythos Preview in April 2026 through Project Glasswing, a restricted program involving partners like AWS, Microsoft, Apple, and CrowdStrike. Anthropic captivated Wall Street and government officials with the unveiling, which excels at identifying security flaws within software. The company said it did not plan to make the model generally available, limiting the rollout to a select group of companies as part of the cybersecurity initiative.

Where Mythos Preview was available to roughly 50 select users, Fable 5 removes that restriction while adding carefully tuned safeguards in high-risk categories. With the launch of Claude Fable 5, Anthropic is honoring its stated "eventual goal" to deploy Mythos-class models at scale.

What makes it different

Fable 5's capabilities exceed those of any model Anthropic has ever made generally available. It is state-of-the-art on nearly all tested benchmarks, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable 5's lead over other models.

The benchmark numbers back that up:

• On SWE-Bench Pro, Fable 5 scored 80.3%, ahead of Claude Opus 4.8 at 69.2%, GPT-5.5 at 58.6%, and Gemini 3.1 Pro at 54.2%.
• On FrontierCode Diamond, Fable 5 scored 29.3%, compared with 13.4% for Claude Opus 4.8 and 5.7% for GPT-5.5.
• On Humanity's Last Exam without tools, Fable 5 scored 59.0%, and 64.5% with tools.
• On HealthBench Professional, a medical reasoning evaluation targeting clinician-level judgment, Fable 5 scored 66.0% versus GPT-5.5 at 51.8%.
• Fable 5 ranks first on the Agent Arena leaderboard, leading by the widest margin ever over Opus 4.8 and GPT-5.5 on confirmed task success rate and praise vs. complaint.

The gap between Fable 5 and its nearest competitors grows as tasks get harder. That is not true of every frontier model, and it is the thing that matters for real deployment.

Built for agents, not just chat

The real story here is not single-turn performance , it is sustained, autonomous work. Run inside an agent harness like Claude Code or Claude Managed Agents, Fable 5 can work for days at a time: planning across stages, delegating to sub-agents, and checking its own work.

Fable 5 is designed for long-running, multi-stage, and asynchronous tasks like complex code refactoring, deep research synthesis, and document-heavy workflows. It can handle tasks previous models could not sustain. It can plan its approach, check progress against the goal, and refine its work as it goes, instead of waiting for the next instruction.

Early partner results are striking. Stripe reported that Fable 5 compressed months of engineering into days , on a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a full team more than two months by hand. One builder burned 700 million tokens in a single day inside Cursor. Another one-shot an entire horror game live on stream.

The safeguard architecture

Getting Mythos-level capability into general hands required solving a hard problem: the same model that makes Fable 5 exceptional at coding also makes it dangerous in the wrong hands. In high-risk areas like cybersecurity, biology, chemistry, and distillation, the model blocks responses and falls back to Claude Opus 4.8.

The fallback mechanism works like this:

1. A separate classifier AI monitors every request in real time.
2. If it detects a query touching cybersecurity, biology, chemistry, or model distillation, it intercepts before Fable 5 responds.
3. The request is silently rerouted to Claude Opus 4.8, and the user is notified.

Anthropic tuned these conservatively , the safeguards will occasionally catch benign requests, but the company says they trigger in less than 5% of sessions on average. The company conducted extensive testing to determine whether users can trick Fable 5 into providing answers it should not, and ran an external bug bounty program with more than 1,000 hours of testing, during which no one found a universal jailbreak for the model.

One practical implication for API integrations: Fable 5 includes safety classifiers that can decline requests, while Mythos 5 does not. If your integration calls Fable 5, plan for three changes: new response handling for refusals, fallback options for retrying on another Claude model, and new billing rules.

A new effort dial replaces temperature

Fable 5 drops the temperature parameter entirely. Instead, it introduces an effort level that controls how deeply the model thinks before responding. The levels are low, medium, high (default), xhigh, and max. On Fable 5, thinking cannot be turned off , effort is the only depth control.

A Hacker News user with 2,093 comments on the launch thread put it clearly: "Fable on 'high' is producing substantially better results than Opus 4.8 on xhigh for me... it 'feels' smarter and doesn't use nearly as many tokens running in circles."

One practical tip from the community: instructions written for old models keep Fable behaving like those old models. Rework your skills and CLAUDE.md files. Let the model use its own judgment first.

Pricing and access

It costs $10 per million input tokens and $50 per million output tokens to access both Fable 5 and Mythos 5. That is exactly double the Opus 4.8 rate, but less than half of the Mythos Preview rate.

Claude Fable 5 is generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. For subscription plans, from launch through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic will remove Fable 5 from those plans, and using it after that will require usage credits.

The model ID for API access is claude-fable-5. Anthropic requires a 30-day data retention period for all Fable 5 and Mythos 5 traffic for safety monitoring purposes.

Where it fits in the bigger picture

Anthropic is effectively introducing a new AI tier above Opus, claiming frontier-level gains in coding, scientific research, cybersecurity, and autonomous work , while using extensive safeguards to enable broader public access. The Fable/Mythos split is a deliberate architectural choice: the same weights, two different access profiles, with the public version trading some capability ceiling in sensitive domains for the ability to ship at all.

Anthropic's latest models come a week after the company announced it filed confidential paperwork for its initial public offering. The timing is not accidental , this is the company's clearest statement yet that its frontier research is ready for production at scale.

For teams building on Claude today, the practical question is straightforward: if your workloads are in coding, knowledge work, finance, or long-horizon agentic tasks, Fable 5 represents a genuine step change. If your work touches cybersecurity or biology, you will hit the fallback layer and should plan accordingly.