Claude Fable 5.1 is the September 1, 2026 update to Anthropic’s Claude Fable 5, the generally available face of its Mythos-class models. Anthropic pitches it as “the world’s most advanced model for coding and knowledge work”, with research capabilities that offer “an early glimpse of how AI models will contribute to scientific progress.” Behind the pitch are three concrete changes: higher capability, a lower bill, and re-tuned safeguards.
It launched together with Claude Mythos 5.1 — the same model with different levels of safeguards, available only through trusted-access programs for cybersecurity and life-sciences work.
What changed from Fable 5
- Price. List price stays $10 / $50, but cache reads fall 75% to $0.25 per million tokens. Because cache reads dominate agent costs, Anthropic estimates typical workloads cost ~25% less and highly agentic ones up to ~45% less.
- Data retention. Enterprise Frontier Safeguards (EFS) store data in cloud infrastructure the customer controls, delivering zero-data-retention-level privacy while still preventing adversarial use. EFS reaches enterprises in phases from later this fall; until then eligible customers can use Fable 5.1 with zero data retention.
- Safeguards. Cybersecurity safeguards now block 60% fewer false positives, partly because Fable 5.1 can be used to discover software vulnerabilities (though not to develop exploits). For biology, an access program built with the US government gives scientists access to Mythos 5.1’s capabilities.
- Preserved thinking. An anti-distillation safeguard introduced with 5.1 stops API users from editing Claude’s prior context to extract its reasoning, applying to API accounts created on or after August 31, 2026.
Benchmarks
Vendor-reported (Anthropic, September 1, 2026), with production safeguards enabled. Where safeguards intervened, Fable models scored zero on OSWorld 2.0, and cyber and biology tasks were completed by Opus 4.8 and Opus 5 respectively — which likely lowers Fable’s scores.
| Benchmark | What it measures | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | Agentic scientific research | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | Agentic terminal coding | 55.8% (Mythos 5.1: 60.9%) | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 | Real-world knowledge work (Elo) | 1853 | 1723 | 1824 | 1711 |
| OSWorld 2.0 | Computer use (partial / strict) | 77.9% / 41.7% | 72.9% / 36.1% | 75.4% / 39.6% | — |
| Humanity’s Last Exam | Reasoning (no tools / tools) | 60.9% / 65.0% | 57.8% / 63.8% | 56.6% / 63.6% | — |
| AutomationBench | Cross-app business workflows | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | Real Cursor coding sessions | 73.4% | 70.5% | 70.0% | 67.2% |
At Low or Medium effort Fable 5.1 matches or beats Fable 5’s results at a much lower cost; it defaults to High effort in Claude Code and Medium in Claude Cowork and on Claude.ai.
The customer anecdotes are unusually concrete: Millennium says Fable 5.1 found the cause of a roughly one-in-a-million crash that none of its engineers or other models had explained in four to five years; Ramp ran a 38-hour unattended machine-learning job in which it diagnosed a prior result as a label artifact and launched six parallel experiments; Cognition moved its Devin Opus 5 traffic to Fable 5.1 on launch day.
Specs and pricing
| Spec | Value |
|---|---|
| API model ID | claude-fable-5-1 |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Thinking | Adaptive, always on |
| Default effort | High in Claude Code; Medium in Cowork and Claude.ai |
| Reliable knowledge cutoff | June 2026 |
| Price (input / output) | $10 / $50 per 1M |
| Cache reads | $0.25 per 1M (was $1 on Fable 5) |
How Fable 5.1 compares
- vs Claude Opus 5.5: Opus 5.5 ($4 / $20, Sept 22) posts higher numbers on Anthropic’s own table, and Anthropic recommends starting there. Fable 5.1 is the pick for the hardest reasoning and long-horizon agentic work.
- vs Claude Fable 5: the same list price with far better science and agentic scores, cheaper caching and fewer false-positive blocks. Practically a replacement.
- vs GPT-6 Astra: Astra ($10 / $50) is priced identically; OpenAI’s own page claims higher scores on Terminal-Bench 4.0 (57.9% vs 55.8%) and Terminal-Bench-Science (64.6% vs 52.6%).
- vs Claude Opus 5: Fable 5.1 leads on Terminal-Bench-Science (52.6% vs 29.0%) and AutomationBench, at twice the price.
Limitations
- Safeguards shape results. Classifiers can reroute cyber and biology requests to older models or zero a benchmark task, so security-adjacent workloads behave differently from the headline numbers.
- Price premium. At 2.5x Opus 5.5’s list price, Fable 5.1 only makes sense where Opus 5.5 at higher effort still falls short.
- Restricted sibling. Mythos 5.1’s capabilities are not available to the public and require approval into a trusted-access program.
For the wider picture, see the 2026 LLM guide.
Best for
- Demanding reasoning and long-horizon agentic work, including multi-day autonomous runs
- Scientific research agents and open-ended problem solving
- Hard debugging and codebase-scale root-cause analysis
- Orchestrator for sub-agents on research and greenfield projects
Pros & cons
Strengths
- Biggest jump over Fable 5 on agentic science — Terminal-Bench-Science 0.1 52.6% vs 24.7% — and on AutomationBench (31.4% vs 17.1%)
- Cheaper to run — cache reads fall 75% to $0.25 per 1M, cutting typical workload costs ~25% and highly agentic ones up to ~45%
- Safeguards retuned — Anthropic reports 60% fewer cyber false positives, and Fable 5.1 can now be used to discover (not exploit) software vulnerabilities
- Customers describe stronger long-run reliability — Millennium traced a crash nobody had explained in 4–5 years; Ramp ran a 38-hour unattended ML job
Limitations
- Still $10 / $50 per 1M — 2.5x Claude Opus 5.5, which Anthropic says performs at Fable 5.1's level on most work
- Safety classifiers can zero out or reroute tasks — Anthropic scored Fable 5.1 as zero where safeguards intervened on OSWorld 2.0, and fallbacks complete cyber tasks on Opus 4.8
- Mythos 5.1 (the same model with fewer safeguards) is limited to trusted-access programs
- Enterprise Frontier Safeguards for full data privacy arrive only in phases from later this fall
Frequently asked questions
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?
They are the same model with different safeguards. Fable 5.1 is generally available; Mythos 5.1 is only available through Anthropic's trusted-access programs and has safeguards designed to support cybersecurity and life-sciences work. Anthropic reports Mythos 5.1 at 60.9% on Terminal-Bench 4.0 vs 55.8% for Fable 5.1.
How much does Claude Fable 5.1 cost?
$10 per 1M input tokens and $50 per 1M output tokens — unchanged from Fable 5 — but cache reads drop 75% to $0.25 per 1M. Anthropic estimates a ~25% lower bill than Fable 5 for typical workloads, and up to ~45% for highly agentic ones.
Should I use Fable 5.1 or Opus 5.5?
Anthropic's guidance: start with Opus 5.5 for most workloads ($4 / $20 per 1M); use Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5.5 at higher effort still fall short. Opus 5.5 beats Fable 5.1 on Anthropic's published benchmark table, but Anthropic says the real-world gap is narrower than the scores suggest.
What are Enterprise Frontier Safeguards?
A system that gives customers privacy equivalent to zero data retention while still blocking adversarial use, by keeping data in cloud infrastructure the customer controls. It rolls out to enterprise customers in phases from later this fall; until then eligible customers can run Fable 5.1 with zero data retention.
What is the Claude API model ID for Fable 5.1?
`claude-fable-5-1`, with a 1M-token context window and up to 128K output tokens. Thinking is adaptive and always on, and the default effort is high in Claude Code and medium in Claude Cowork and on Claude.ai.