Claude Fable 5

Claude Fable 5 is Anthropic's June 2026 public Mythos-class model. It is positioned as the broadly usable version of the Mythos capability tier: stronger than Claude Opus 4.8 on long, complex coding and knowledge-work tasks, but with safeguards that route high-risk requests to the safer Opus 4.8 model.

Core profile

Claude Fable 5 is built for long-horizon work: agentic coding, multi-step research, legal and financial analysis, computer use, and workflows that require the model to hold state across large contexts and many tool calls.

Anthropic prices Fable 5 at USD 10 per million input tokens and USD 50 per million output tokens. That is double Claude Opus 4.8's regular price, but below the earlier Mythos Preview pricing. Anthropic also applies its prompt caching discounts, which can materially lower input cost for repeated long-context workflows.

The important access caveat is that Fable 5 is not simply "Opus, but better" on every plan forever. At launch, Anthropic included it on Claude Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026, after which usage credits apply unless Anthropic extends included access based on capacity.

Why it matters

Fable 5 is the first Mythos-class Claude model Anthropic made generally available. Claude Mythos Preview had been restricted through Project Glasswing because of its advanced cybersecurity and scientific-risk capabilities. Fable 5 uses the same broad capability tier but adds hard safety routing: when a request touches sensitive areas such as cybersecurity, biology, chemistry, or model distillation, Anthropic says the system falls back to Claude Opus 4.8.

That split matters because it separates general frontier capability from restricted high-risk capability. For most coding, knowledge work, document analysis, and enterprise agent tasks, users get the stronger Fable behavior. For the domains that made Mythos dangerous, users get a safer fallback unless they are approved for Mythos 5 through controlled access.

Benchmark profile

Anthropic's launch materials presented Fable 5 as a major jump over Opus 4.8 on difficult work:

• SWE-bench Pro: 80.3%, compared with Opus 4.8 at 69.2%
• FrontierCode Diamond: 29.3%, compared with Opus 4.8 at 13.4%
• GDPval-AA Elo: 1932, ahead of Opus 4.8 at 1890
• OSWorld-Verified: 85.0% on computer-use tasks
• Blueprint-Bench 2: 38.6% on spatial reasoning
• Legal Agent Benchmark: 13.3%, ahead of GPT-5.5 and Gemini 3.1 Pro in Anthropic's comparison table

Some launch benchmarks are starred because they reflect Mythos 5 rather than the public Fable 5 behavior. This is especially important for cybersecurity and biology evaluations. The model family may have the capability, but the public model can route those requests away from Fable 5.

The DeepSWE caveat

Fable 5's headline coding number is on SWE-bench Pro, a benchmark that became more controversial after DeepSWE documented verifier issues and a git-history loophole affecting some Claude-family runs. As of Fable 5's launch, there was no public DeepSWE score for Fable 5.

That does not mean Fable 5 is weak. It means the 80.3% SWE-bench Pro result should be read as a strong but incomplete signal until independent evaluations test the model on stricter coding benchmarks and real production harnesses.

What it is good at

Fable 5 is best used as an escalation model for tasks where cheaper models tend to fail quietly: large codebase changes, multi-file debugging, complex refactors, agentic coding runs, legal or financial document reasoning, long analytical workflows, and research tasks that require sustained coherence over many steps.

Its value is strongest when the task is long enough that better reasoning can reduce retries, tool mistakes, and back-and-forth. For short chat, routine autocomplete, simple summarization, or high-volume low-complexity work, Opus 4.8 or cheaper models may be more cost-effective.

Tradeoffs

The tradeoffs are cost, access, and safety routing. Fable 5 is expensive at USD 50 per million output tokens, which makes it hard to justify for every request in a production system. Its broad-plan inclusion also started as time-limited access rather than a permanent promise. And for high-risk domains, the public model may not expose the Mythos-level behavior that appears in starred benchmark rows.

In practice, Fable 5 is not the everyday default for most teams. It is the model you route to when the task is hard enough that paying more for fewer failures is cheaper than debugging a weaker model's mistake later.