How Claude Fable 5 ranks on benchmarks
Claude Fable 5 led CursorBench 3.1 at launch and remains the reference point for newer frontier coding models. Here are the numbers and the caveat that matters.
•3 min read•Written by
Agents Directory@agentsdir
Anthropic released Claude Fable 5 on June 9. It's the company's first Mythos-class model, priced at $10 in / $50 out per million tokens with a 1M-token context window and built for long-running autonomous work. Here is where it lands, sourced from Anthropic's announcement and the independent CursorBench leaderboard.
CursorBench 3.1
CursorBench evaluates models on ambiguous, multi-file tasks taken from real Cursor sessions. It's the closest thing we have to a production agentic-coding benchmark.
- Fable 5 high (default): 70.6% at $10.81 per task, more than 7 points clear of every other default configuration.
- Fable 5 Max: 72.9%, the top score on the whole leaderboard.
- Next-best defaults: Cursor's Composer 2.5 at 63.2% ($0.55 per task, the value outlier), GPT-5.5 high at 62.6%, Claude Opus 4.8 high at 58.4%.
Artificial Analysis Intelligence Index
Artificial Analysis publishes a composite 0-100 intelligence score that blends knowledge, reasoning, math, coding, and agentic evaluations. It is the most widely cited all-up benchmark outside vendor tables.
- Fable 5 (default): 59.9, the top score in our catalog, 1 point above GPT-5.6 Sol max (58.9).
- Next in this cut: Claude Opus 4.8 (55.7), GPT-5.5 xhigh (54.8), Grok 4.5 high (53.8), Claude Opus 4.7 (53.5), and Claude Sonnet 5 max (53.4).
What changed after launch
Fable 5 remains the leader in this catalog cut, but the frontier around it has tightened. Artificial Analysis measured GPT-5.6 Sol max at 58.9 on its current Intelligence Index, only 1 point behind Fable 5. Grok 4.5 high reached 53.8.
The coding picture now depends on the harness. OpenAI reports that GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1, and Artificial Analysis places Sol ahead on its Coding Agent Index in the Codex harness. Fable 5 still leads the CursorBench results above and remains ahead of Grok 4.5 in both SpaceXAI's DeepSWE tables and its Terminal-Bench 2.1 comparison. The evidence now supports benchmark-specific leaders, not one universal winner.
Anthropic's reported numbers
From the announcement, against Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro, with the best score per row highlighted:
| Claude Mythos 5 / Fable 5 | Claude Mythos Preview | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | |
|---|---|---|---|---|---|
Agentic coding SWE-Bench Pro | 80.3% | 77.8% | 69.2% | 58.6% | 54.2% |
Agentic coding FrontierCode (Diamond), xhigh | 29.3% | — | 13.4% | 5.7% | — |
Agentic coding Terminal-Bench 2.1 | 88.0%* | — | 82.7% | 83.4% Codex CLI | 70.7% Gemini CLI |
Knowledge work GDPval-AA | 1932 | — | 1890 | 1769 | 1314 |
Knowledge work vision GDP.pdf, no tools | 29.8% | — | 22.5% | 24.9% | 16.7% |
Spatial reasoning Blueprint-Bench 2 | 38.6% | — | 14.5% | 36.2% | 26.5% |
Tool use AutomationBench | 17.4% | — | 15.5% | 12.9% | 9.6% |
Computer use OSWorld-Verified | 85.0% | 85.4% | 83.4% | 78.7% | 76.2% |
Legal Legal Agent Benchmark | 13.3% | — | 10.4% | 2.1% | 0.0% |
Multidisciplinary reasoning Humanity's Last Exam, no tools | 59.0%* | 56.8% | 49.8% | 41.4% | 44.4% |
Multidisciplinary reasoning Humanity's Last Exam, with tools | 64.5%* | 64.7% | 57.9% | 52.2% | 51.4% |
Biology BioMysteryBench, hard | 46.1%* | 29.6% | 40.0% | — | — |
Biology BioMysteryBench, human solved | 83.9%* | 82.6% | 80.4% | — | — |
Cybersecurity ExploitBench (Cap) | 78.0%* | 69.0% | 40.0% | 34.0% | — |
Health HealthBench Professional | 66.0%* | 64.7% | 56.9% | 51.8% | — |
Anthropic reports the higher score of Claude Mythos 5 and Claude Fable 5; the two land within 1-3 points of each other except on starred (*) benchmarks. See the Mythos caveat below.
The Mythos caveat
Anthropic reports the higher score of two models: Claude Mythos 5 (the identical model with safeguards lifted, restricted to vetted researchers) and the generally available Fable 5. The two land within 1-3 points of each other except on starred (*) benchmarks, where Fable 5's safeguards redirect cybersecurity and biology queries to Opus 4.8 (under 5% of sessions). On those, Fable 5's effective score sits closer to Opus 4.8.
Bottom line
At twice Opus 4.8's price, Fable 5 is not the default for everything. It remains the strongest model on the CursorBench results above, while GPT-5.6 Sol now leads other agentic coding evaluations and Grok 4.5 offers a lower-cost alternative. The CursorBench cost curve still shows Fable's premium buying capability, not just tokens. Pricing, host availability, and sources are on the Claude Fable 5 model page.
Sources
- Anthropic: Introducing Claude Fable 5 and Claude Mythos 5 for the announcement, pricing, and the reported benchmark table
- CursorBench leaderboard for independent agentic-coding scores and per-task cost
- CursorBench 3.1 on Agents Directory for the live leaderboard we keep updated
- Artificial Analysis Intelligence Index for the composite intelligence methodology
- Artificial Analysis GPT-5.6 results for current Sol, Terra, and Luna Intelligence Index scores
- OpenAI: GPT-5.6 for the model family's release details and reported agentic benchmark positioning
- SpaceXAI: Introducing Grok 4.5 for its reported coding comparisons with Fable 5
- Intelligence Index on Agents Directory for the live leaderboard we keep updated
- Claude Fable 5 model page for pricing, context window, and host availability