Claude Fable 5: What Developers Need to Know

Anthropic Just Shipped Something Different

On June 9, 2026, Anthropic did something it has never done before. It took a model from its top-secret Mythos tier — the class of models previously accessible only to vetted cybersecurity partners and select biology researchers under Project Glasswing — and shipped a production-safe version of it to the public.

That model is Claude Fable 5. It does not sit inside the Opus family. It sits above it. If you have been watching frontier model releases inch forward by half a benchmark point at a time, Fable 5 is a different kind of release.

This post covers what Fable 5 actually is, what the benchmarks mean for developers, how it is priced, how to call it, and the one architectural decision Anthropic made that every developer building on it needs to understand before shipping.

🎯 Quick Answer (30-Second Read)

What it is: The first publicly available Mythos-class model from Anthropic — above Opus in capability tier
Standout benchmark: 80.3% on SWE-Bench Pro — 11 points ahead of Opus 4.8, 21.7 points ahead of GPT-5.5
API model string: claude-fable-5
Pricing: $10 per million input tokens / $50 per million output tokens (batch: $5/$25)
Key caveat: On cybersecurity, biology, chemistry, and distillation queries, Fable 5 falls back to Opus 4.8 silently — in under 5% of sessions
Availability: Free on Pro, Max, Team, and Enterprise until June 22 — usage credits required after that until capacity expands

What Fable 5 Actually Is

Fable 5 is not a new model family — it is the public-safe version of Claude Mythos 5, the same underlying model that has been running inside Project Glasswing since April 2026. Anthropic wrapped it in safety classifiers that intercept a specific set of sensitive query categories and route them to Opus 4.8 instead. Outside those categories, Fable 5 and Mythos 5 are effectively the same model.

This matters for how you think about the product. You are not getting a slightly better Opus. You are getting Mythos-class capability on the full range of tasks that do not touch the restricted domains — with a fallback layer handling the rest.

The restricted sibling, Claude Mythos 5 (API string: claude-mythos-5), remains limited to approved Project Glasswing partners and selected infrastructure providers. It is not generally available.

The Fallback Architecture

Anthropic is explicit about this and it is worth understanding before you build on it. When a query trips one of Fable 5's safety classifiers — covering cybersecurity, biology, chemistry, model distillation, and frontier AI development — the response is handled by Claude Opus 4.8 instead, and the API tells you it happened. Anthropic says this triggers in fewer than 5% of sessions.

For most developer workflows — coding, knowledge work, document reasoning, agentic tasks — you will never see a fallback. For workflows that touch any of those restricted domains, you need to account for it.

The Benchmarks That Actually Matter

Benchmark releases are usually table stakes noise. The numbers on Fable 5 are not.

SWE-Bench Pro (agentic coding):

Fable 5: 80.3%
Opus 4.8: 69.2%
GPT-5.5: 58.6%
Gemini 3.1 Pro: 54.2%

The gap between Fable 5 and Opus 4.8 is larger than the gap between Opus 4.8 and Gemini 3.1 Pro. That is not a marginal improvement — that is a tier jump.

FrontierCode Diamond (hardest coding eval):

Fable 5: 29.3%
Opus 4.8: 13.4%

Other benchmarks (vs Opus 4.8):

GDPval-AA (knowledge work): 1932 vs 1890
OSWorld-Verified (computer use): same underlying model
Legal benchmark: +13.3 points
Health: 66.0% (Mythos 5 score — Fable 5 lands closer to Opus 4.8 due to fallbacks)

The pattern across all of these is consistent: the largest gains are on the longest, hardest tasks. This is exactly where agentic workloads live. Cursor CEO Michael Truell confirmed this in real-world testing: Fable 5 opened up a class of long-horizon problems that were previously out of reach.

How to Use Fable 5 in Production

Calling Fable 5 works identically to any other Claude model through the Messages API:

const response = await fetch("https://api.anthropic.com/v1/messages", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "x-api-key": process.env.ANTHROPIC_API_KEY,
    "anthropic-version": "2023-06-01"
  },
  body: JSON.stringify({
    model: "claude-fable-5",
    max_tokens: 8096,
    messages: [
      { role: "user", content: "Your prompt here" }
    ]
  })
});

The model string is claude-fable-5. No other changes required if you are already calling Opus 4.8.

Cost Control Is Not Optional

Fable 5 is expensive — $10 per million input tokens, $50 per million output tokens. On long agentic tasks it will burn tokens fast. Route it through a gateway if you are running it at volume, cache aggressively, and reserve it for tasks that actually justify the cost.

Opus 4.8 at its Fast mode pricing is still the right call for routine, latency-sensitive, or cost-sensitive work. Fable 5 is for the hard problems.

Fable 5 vs Opus 4.8 — When to Use Which

Task	Use Fable 5	Use Opus 4.8
Long-horizon agentic coding	✓
Complex codebase migrations	✓
Multi-step document reasoning	✓
Routine API calls and chat		✓
Cost-sensitive high-volume workloads		✓
Latency-sensitive pipelines		✓
Cybersecurity / bio / chemistry queries	Fallback to 4.8	✓

My Take

The reason Fable 5 is a bigger deal than another benchmark release is that the gains are concentrated exactly where AI products have been bottlenecked: long-horizon tasks, complex reasoning across large contexts, and agentic coding on real codebases. Most frontier model releases give you marginal gains on standard benchmarks. An 11-point jump on SWE-Bench Pro while the next-best model is your own previous flagship is not marginal — that is what a capability tier change looks like. The best outcome here is that developers start building workflows around genuinely long-horizon tasks they previously could not trust to a model. The worst outcome is teams burning through credits on Fable 5 for tasks that Opus 4.8 handles perfectly well, then concluding it is not worth the cost. The fallback architecture is the honest version of what responsible frontier deployment looks like in practice — and every lab will be forced toward something like it as models get more capable. Where this goes: Mythos 5 becoming more broadly available as Anthropic validates the safety classifiers at scale is the obvious next step. Whether the restricted domains stay restricted that long is the real question.

Real Developer Use Case

Stripe was among the early testers. Their engineering team ran Fable 5 against a 50-million-line codebase migration that had previously required manual coordination across multiple teams. Fable 5 completed it in a day.

That is not a benchmark number — that is a workflow that was previously not viable as an AI-assisted task now becoming routine. For developer teams running large-scale refactors, dependency upgrades, or cross-codebase migrations, the practical ceiling has moved.

Frequently Asked Questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, launched June 9, 2026. It is the same underlying model as Claude Mythos 5 — the restricted frontier model that has been running in Project Glasswing since April 2026 — wrapped in safety classifiers for general use. It sits above the Opus family in Anthropic's model tier hierarchy.

What is the Claude Fable 5 API model string?

The API model string is claude-fable-5. The restricted sibling is claude-mythos-5, which is not publicly available. Calling Fable 5 via the Messages API requires no other changes from a standard Claude API setup.

How much does Claude Fable 5 cost?

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. Batch pricing is $5 per million input tokens and $25 per million output tokens. It counts as 2x usage on Claude.ai subscription plans.

What is the fallback to Opus 4.8 and when does it trigger?

When a Fable 5 query trips a safety classifier — covering cybersecurity, biology, chemistry, model distillation, and frontier AI development — the response is automatically handled by Claude Opus 4.8 instead. The API indicates when a fallback has occurred. Anthropic says this triggers in fewer than 5% of sessions.

Is Claude Fable 5 better than GPT-5.5?

On agentic coding benchmarks, yes — significantly. Fable 5 scores 80.3% on SWE-Bench Pro versus GPT-5.5's 58.6%. GPT-5.5 leads on Terminal-Bench 2.1 through its own Codex CLI harness. For long-horizon coding and complex knowledge work, Fable 5 has the stronger benchmark position. GPT-5.5 costs roughly half as much, which matters for cost-sensitive workloads.

Conclusion

Claude Fable 5 is the most capable model Anthropic has made publicly available. The benchmark gains are real and concentrated on the workloads — long-horizon agentic tasks, complex coding, deep document reasoning — where developers have been hitting the ceiling of what frontier models could reliably handle.

Use it for hard, long problems where the quality gap justifies the cost. Use Opus 4.8 for everything else. Understand the fallback architecture before you build a workflow that touches the restricted domains.