Claude Fable 5.1 Deep Dive: Making Long-Horizon Agentic AI the Main Battleground

*Research date: 2026-09-02. Subject: Anthropic Claude Fable 5.1 / Mythos 5.1 (released 2026-09-01). Primary source: Anthropic's official benchmark comparison table (7 benchmarks × 4 models, with 4 footnotes). Corroborating sources: Laude Institute Terminal-Bench-Science 0.1 leaderboard, Zapier AutomationBench, Cursor CursorBench 3.2, Artificial Analysis GDPval-AA v2 and AutomationBench-AA, OSWorld 2.0 paper, CAIS/Scale AI HLE.*

Data grading: CONFIRMED = official leaderboard/paper verifiable | VENDOR = vendor-run, not independently reproduced | RUMOR = customer testimonies / third-party retelling.

Key points

• Headline numbers: Terminal-Bench-Science 0.1: 52.6% (2.13× Fable 5's 24.7%); HLE no-tools: 60.9% (first public model above 60%); GDPval-AA v2 Elo 1853 — the only loss, to Opus 5's 1861, within sampling error; cache-read price cut 75% ($1.00 → $0.25/M tokens); up to ~45% cost reduction on highly agentic workloads.
• One-line verdict: Fable 5.1 is not a "Fable 5 + 0.1" patch. It is the first time Anthropic places "completing an entire task autonomously" at the center of the product claim — first place on six of seven agentic benchmarks.

Timeline: three generations in three months

| Date | Event | Gap | Positioning | |---|---|---|---| | ~2026-05-20 | Fable 5 released | — | First "Fable" tier, above Opus | | 2026-07-24 | Opus 5 released (1M context default + adaptive thinking) | ~65 days | GDPval-AA v2 record 1861 | | 2026-09-01 | Fable 5.1 + Mythos 5.1 | 39 days after Opus 5 | First on 6/7 benchmarks |

The flagship shelf life is compressing: Fable 5 was replaced by its own successor after just three months. Model selection is now a quarterly, not annual, decision.

Twins: one base model, two safeguard levels

• Fable 5.1 (claude-fable-5-1) — open to everyone; available on AWS / Google Cloud / Azure.
• Mythos 5.1 — same base model, looser safeguards; restricted to vetted cybersecurity professionals and life-science researchers via the Cyber and Life Sciences Verification Programmes. Currently limited to some US organizations. Anthropic calls it its strongest cybersecurity model to date.

Pricing: base rates unchanged, the dominant cost item slashed

| Item | Fable 5 | Fable 5.1 | Change | |---|---|---|---| | Input / per M tokens | $10.00 | $10.00 | — | | Output / per M tokens | $50.00 | $50.00 | — | | Cache read | $1.00 | $0.25 | −75% | | 5-min cache write | — | $12.50 | — | | 1-hour cache write | — | $20.00 | — |

Cache reads are a *multiple* of input price: other Claude models sit at 0.1×, Fable 5.1 at 0.025×. Since every agent loop re-reads system prompts, history, and tool results, cache reads dominate long-task bills. Official guidance: ~25% total cost reduction on typical workloads, up to ~45% on highly agentic workloads. Context: per payment platform Ramp, Fable 5 accounted for only ~11% of enterprise AI tool spend two months after launch — cheaper Opus 5 outsold it. Fable's bottleneck was never capability; it was the bill.

The seven benchmarks

1. Agentic science — Terminal-Bench-Science 0.1 [CONFIRMED]

70 tasks drawn from working scientists' pipelines (broken FASTA files, half-written simulation docs, data cleaning), graded by deterministic pytest across life science (19), physical science (17), earth science (8), math (17), engineering (9). Selection was brutal: 920 proposals → 464 approved → 386 PRs → 70 tasks, a 7.6% survival rate.

| System | Solve rate | |---|---| | Fable 5.1 (Anthropic harness) | 52.6% (+27.9pp vs Fable 5, 2.13×) | | Opus 5 (Claude Code) | 30.0% (previous Laude leaderboard record) | | Fable 5 | 24.7% (21.4% per Laude's own leaderboard) | | GPT-5.6 Sol (Codex) | 22.4% | | Opus 4.8 | 10.5% | | GLM 5.3 | 8.1% (best open-weight) |

This is the real headline: the public record for autonomous scientific work jumps from 30.0% to 52.6%. Two deductions: (1) 52.6% is VENDOR-graded — Anthropic's harness differs from Laude's by ~3.3pp on the same model; (2) footnote 1 reveals biology tasks blocked by safeguards were completed by Opus 5 — and life science is the largest domain (19 tasks), so 52.6% is likely underestimated.

2. Agentic coding — Terminal-Bench 4.0 [VENDOR]

Version 4.0 recalibrated time/CPU/memory quotas, dropped 8 tasks, fixed 19 — scores are not comparable to earlier versions. Fable 5.1: 55.8%, vs Mythos 5.1 60.9%, Opus 5 52.3%, Fable 5 42.0%, GPT-5.6 Sol 37.3% (+18.5pp, 1.50×). Note: Mythos 5.1's 60.9% comes from a third-party tracker and exactly matches Fable 5.1's HLE no-tools score — flagged as needing verification.

3. Professional knowledge work — GDPval-AA v2 [CONFIRMED]

OpenAI's task set (44 occupations, 9 industries), Artificial Analysis's agentic version: 220 tasks with shell/web access, blind pairwise LLM judging into Elo, human baseline anchored at 1000.

| System | Elo | |---|---| | Opus 5 | 1861 | | Fable 5.1 | 1853 (−8; pairwise win rate 51.2% — statistically indistinguishable) | | Fable 5 | 1747 | | GPT-5.6 Sol | 1738 |

Fable 5.1 effectively tied Opus 5. Three caveats: the judge is an LLM, not a human; Elo swings heavily by effort tier (Opus 5 scored 1852/1819/1735 at max/xhigh/high — a 117-point spread, an order of magnitude larger than the 8-point gap); sources disagree on Opus 5's exact value (1861/1852/1853). Anthropic's own guidance: start with Opus 5; move to Fable 5.1 when needing stronger reasoning plus long-horizon agentic behavior. Two different knives, not a clean sweep.

4. Computer use — OSWorld 2.0 [CONFIRMED]

108 end-to-end workflows across 31 self-hosted sites and desktop apps; tasks take humans ~1.6 hours, agents ~318 tool calls. Dual metrics: *partial* (partial credit, ~27.25 checkpoints/task) and *strict* (binary completion).

| System | partial | strict | strict/partial conversion | |---|---|---|---| | Fable 5.1 | 77.9% | 41.7% | 53.5% | | Fable 5 | 72.9% | 36.1% | 49.5% | | Opus 5 | 70.6% | – | – | | GPT-5.6 Sol | 62.6% | – | – | | Opus 4.8 | 54.8% | 20.6% | 37.6% | | GPT-5.5 | 47.5% | 13.0% | 27.4% |

The sharpest derived metric: strict/partial conversion — "of tasks already 70% done, how many actually finish?" Fable 5.1 improves it to 53.5%: it doesn't just do more, it closes out. But 41.7% strict means 58.3% of long-horizon computer tasks still fail; the OSWorld paper's own line — "current agents remain far from professional-grade computer use" — has narrowed but not closed.

5. Multidisciplinary reasoning — Humanity's Last Exam [CONFIRMED in part]

2,500 public questions across 100+ subjects; released with GPT-4o at just 2.7% in Jan 2025.

| System | No tools | With tools | Tool gain | |---|---|---|---| | Fable 5.1 | 60.9% | 65.0% | +4.1pp | | Fable 5 | 53.7% | 57.6% | +3.9pp | | Opus 5 | 50.2% | 54.9% | +4.7pp | | GPT-5.6 Sol | 47.4% | 52.1% | +4.7pp |

60.9% is the first time a publicly available model clears 60% without tools. The hardest, most overlooked signal: tool gains are nearly identical across all four models. Fable 5.1's edge is pure reasoning, not tool chaining — more credible than the with-tools number. Human baselines: ~90% for domain experts in-domain, 60–70% cross-domain. Fable 5.1 (65.0% with tools) sits at the low end of "cross-domain expert." Remember HLE's known flaw: a July 2025 FutureHouse review found 29% of chemistry/biology answers contradicted the literature; HLE-Rolling revision is ongoing.

6. Business process automation — AutomationBench [VENDOR]

Zapier's benchmark spins up a miniature simulated company (CRM, inbox, calendar, tickets; 47 SaaS apps, ~500 API endpoints, ~12,000 assertions) and judges the end state of the world, not the agent's words. Fable 5.1: 31.4% vs Fable 5's 17.1% (+14.3pp, 1.84× — second-largest generational jump). On Zapier's own leaderboard, Fable 5 (17.4%) had barely edged Opus 4.8 (17.2%) and lost to GPT-5.6 Sol (18.1%); Fable 5.1's jump reopens the gap. But cross-vendor comparison here largely fails: Zapier uses a private held-out set, Anthropic its own harness, and Artificial Analysis's AutomationBench-AA gave Fable 5 48.6% — the same model differs up to 3× across harnesses. Also notable: on AutomationBench-AA, every tested model violated business guardrails (0.46–1.26 violations per task); Anthropic reported no guardrail data this time.

7. In-IDE coding — CursorBench 3.2.0 [VENDOR]

Tasks from real Cursor sessions; v3.2 adds instruction following and advanced tool use.

| System | Score | Cost/task | Tokens/task | Steps | |---|---|---|---|---| | Fable 5.1 | 73.4% | n/a | n/a | n/a | | Fable 5 (Max) | 70.5% | $17.32 | 103,525 | 72 | | Opus 5 (Max) | 70.0% | $8.23 | 61,838 | 78 | | GPT-5.6 Sol (Max) | 67.2% | $5.69 | 28,320 | 48 | | Grok 4.6 | 70.8% | – | – | – |

This "win" deserves the biggest discount: +2.9pp over Fable 5, with Anthropic's own footnote stating ±3.5pp error at n=100 — statistically indistinguishable. The real story is the cost column: Fable 5 cost 3× GPT-5.6 Sol per task and burned 3.7× the tokens for +3.3pp. That's exactly what the 75% cache-read cut targets. Estimated Fable 5.1 (Max) cost: roughly $9.5–13/task — still well above GPT-5.6 Sol's $5.69.

Scorecard at a glance

| Benchmark | vs Fable 5 | vs Opus 5 | vs GPT-5.6 Sol | Verdict | |---|---|---|---|---| | Terminal-Bench-Science 0.1 | +27.9pp (2.13×) | n/a | +30.2pp | Decisive win | | Terminal-Bench 4.0 | +13.8pp | +3.5pp | +18.5pp | Win | | GDPval-AA v2 | +106 Elo | −8 Elo | +115 Elo | Tie with Opus 5 | | OSWorld 2.0 partial | +5.0pp | +7.3pp | +15.3pp | Win | | OSWorld 2.0 strict | +5.6pp | n/a | n/a | Win (leader, no comparison) | | HLE no-tools | +7.2pp | +10.7pp | +13.5pp | Decisive win | | HLE with tools | +7.4pp | +10.1pp | +12.9pp | Decisive win | | AutomationBench | +14.3pp (1.84×) | n/a | n/a | Big generational gain, weak cross-vendor basis | | CursorBench 3.2.0 | +2.9pp | +3.4pp | +6.2pp | Within error bars |

Bottom line

Fable 5.1's substance is the leap in autonomous scientific work (24.7% → 52.6%), long-horizon task completion (strict OSWorld conversion up to 53.5%), and pure reasoning (first public 60%+ on HLE no-tools) — combined with pricing that finally attacks its real weakness: cost per completed agentic task. Its limits: only 41.7% strict completion on computer use, vendor-run harnesses on three of seven benchmarks, a tie rather than a win on professional knowledge work, and per-task economics still well behind GPT-5.6 Sol. The release cadence — three flagship-tier models in three months — tells its own story about where the competitive frontier now sits.