Anthropic: Claude Fable 5
Mythos-class frontier model from Anthropic with a 1,000,000-token context window, 95% on SWE-bench Verified, and a 64.9 Intelligence Index. Safety classifiers route sensitive queries to Claude Opus 4.8, and access has been restored after a temporary June 2026 suspension.
Specifications
| Attribute | Value |
|---|---|
| Lab | Anthropic |
| Tags | Agentic Coding Intelligent Reasoning Long Context |
| Release Date | 2026-06 |
| Context Window | 1,000,000 tokens |
| Input Price / 1M | $10.00 |
| Output Price / 1M | $50.00 |
| Input Modalities | Image, Text |
| Output Modalities | Text |
Strengths
- #1 Intelligence Index at 64.9
- 1M-token context window
- 95% on SWE-bench Verified
- Strong agentic and coding performance
- Vision and scientific research capable
Weaknesses
- Safety fallback reduces capability on sensitive queries
- Access may still be capacity-constrained after the June 2026 suspension
- High cost: $10/$50 per million tokens
- Mandatory 30-day data retention
Best For
- Autonomous agentic workflows
- Long-horizon software engineering
- Complex knowledge work and research
- Multi-step tool-use with verifiable chain-of-thought
In Depth: Claude Fable 5
Benchmark Performance
Claude Fable 5 leads LMRank with a 9.9/10 score and the #1 Intelligence Index of 64.9, outperforming the next-best model by 7 points on AA-Omniscience and 7 points on Humanity's Last Exam over Claude Opus 4.8.
Independent evaluations show Fable 5 scoring 95.0% on SWE-bench Verified (official), 80.0% on SWE-bench Pro, and 29.3% on FrontierCode Diamond. GPQA Diamond reaches ~93% when non-fallback, though safety triggers drop effective performance to Opus 4.8 levels on sensitive domains. Terminal-Bench Hard scores ~84.3% with a ~21% fallback rate.
Pricing & Value
Claude Fable 5 costs $10.00 per million input tokens and $50.00 per million output tokens. Cached reads are $1.00 per million tokens, with cache writes at $12.50.
At $50 per million output tokens, Fable 5's cost-per-performance is competitive among frontier models: its Intelligence Index of 64.9 yields ~1.30 points per dollar of output, compared to Gemini 3.1 Pro's 57.9 at $35 output (~1.65 points per dollar). For coding-heavy workloads, the 95% SWE-bench Verified score justifies the premium over Opus 4.8 ($15/$60).
Who Should Use This
Claude Fable 5 is built for enterprises pushing autonomous workflows, researchers tackling long-context analysis, and engineering teams automating multi-step code tasks.
- Autonomous agent builders: top reasoning but safety fallback triggers on sensitive domains
- Enterprise engineers: excellent long-horizon coding, high cost per token
- Scientific researchers: strong vision and analysis, 30-day data retention requirement
- Not for: applications requiring unfiltered access to cybersecurity or biology knowledge
Release & Version History
Claude Fable 5 launched June 9, 2026 as Anthropic's first Mythos-tier model, with access restored after a short demand-driven suspension three days later.
Fable 5 shares underlying weights with Claude Mythos 5, which is restricted to Glasswing partners for US government cyberdefense. It adds safety classifiers that fallback to Claude Opus 4.8 on sensitive queries. The model succeeded Claude Opus 4.8 and follows Anthropic's pattern of tiered safety deployment across public and restricted releases.