SWE-bench Pro 80.3%. GPT-5.5 stalled at 58.6% — and Fable 5 blew right past it. Anthropic's first publicly available Mythos-class model is here. And the price is exactly double.
What does "21 points ahead of GPT-5.5" actually mean?
Claude Fable 5 is the first publicly available "Mythos-class" model from Anthropic. Under the hood, it shares the same base model as Claude Mythos 5 — with safety filters layered on top for general release.
The benchmark numbers are hard to ignore, especially in agentic coding:
Real-world enterprise evaluations back it up. Hex confirmed "90% on core analytics benchmarks," and Genspark reported Fable 5 "outperformed all competing models on UI design and game coding evaluations." Base44 noted superior capability for building complete applications in single attempts.
That said, the gap is most pronounced in agentic coding. On general knowledge benchmarks, GPT-5.5 remains competitive.
Here's the catch: it blocks 8–9% of requests
Fable 5 includes three safety classifiers. When cybersecurity, biology/chemistry queries, or model distillation attempts are detected, it automatically falls back to Opus 4.8. The issue is that these filters are broader than you'd expect.
Heads up for science and medical users
Reports include a medical physicist getting blocked for using the word "nuclear," MRI segmentation tasks being flagged, and malaria research queries being refused. Independent testing puts Fable 5's overall refusal rate at roughly 8–9%, concentrated in scientific and technical queries.
The 30-day data retention policy is another variable. Even organizations with existing zero-retention agreements now have Fable 5 and Mythos 5 traffic retained for 30 days automatically. If your workflow touches sensitive data, review your security policy before deploying.
The 2x price is only justified for specific tasks
Fable 5 API pricing: $10 input / $50 output per million tokens — exactly double Opus 4.8 ($5/$25). Route everything through it and costs double. The key is knowing when to use it and when not to.
| Task type | Opus 4.8 / Sonnet | Fable 5 |
|---|---|---|
| Summaries, translations, editing | Works fine | Overkill |
| Everyday code writing and bug fixes | Works fine | Overkill |
| Long-running autonomous coding agents | Hits limits | Worth it |
| Large codebase refactoring | Hits limits | Worth it |
| Multi-step research automation | Hits limits | Worth it |
| Complex architecture design and review | Hits limits | Worth it |
The core principle: "The 2x price is justified when it handles work that would otherwise take a human several days." For everything else, the relay pattern — routing to Opus 4.8 or Sonnet — maintains quality while cutting costs in half.
How to get started with Fable 5
- Check your plan or API access
Available on Pro, Max, Team, and Enterprise subscriptions, or via API (model ID: claude-fable-5-20260609). Amazon Bedrock, Google Cloud, and Microsoft Foundry are also supported. - Set up the relay pattern
Send design, review, and architecture work to Fable 5; route repetitive implementation, translations, and summaries to Opus 4.8 or Sonnet 4.6. This typically cuts total project costs by 40–60%. - Enable prompt caching
Apply caching to repeated system prompts or large codebases to cut input costs by 60–80%. The Batch API adds another 50% discount on top. - Test safety filters first
If you work in science or medicine, test your actual workflows before full integration. Claude Code v2.1.170+ lets you select the model with the /model command. - Review the data retention policy
30-day traffic retention applies to all Fable 5 usage. For workflows with sensitive data, review your data processing agreements before deploying.
Want to go deeper?
Claude Fable 5 and Mythos 5 Official Announcement Full breakdown from Anthropic: capabilities, benchmarks, safety architecture, and Project Glasswing. anthropic.com
Claude Fable 5: powerful, expensive, and heavily filtered The Decoder's independent benchmark analysis and real user cases. Good for understanding the actual impact of the 8–9% refusal rate. the-decoder.com
Claude Fable 5 Launch Guide 2026 Independent benchmarks including SWE-bench Pro 80% with detailed GPT-5.5 comparison. codersera.com
Claude Fable 5 API Pricing & Benchmarks Live pricing, throughput, and uptime data on OpenRouter. Useful for API routing optimization. openrouter.ai
Claude Mythos Pricing Guide Token cost breakdowns for Fable 5 and Mythos 5 with ROI calculations vs Opus 4.8. cloudzero.com
