Claude Fable 5

The Agentic Coding Leader, Measured

Anthropic’s Fable 5 does not win the most benchmarks - it wins the ones that are hardest to fake. It tops SWE-Bench Pro (80.3%), the contamination-resistant split where memorizing public test sets does not help; it tops Terminal-Bench 2.1 (88.0%), where models must actually run, debug, and repair live environments; and it tops GDPVal-AA (1932 Elo), the knowledge-work evaluation built from real professional tasks. Three number-one results, all in categories where the evaluation is the job.

The headline number most teams care about: 84.5 on SWE-bench Verified. Read our SWE-bench Pro guide for why that Verified-versus-Pro distinction matters when comparing vendors’ claims.

Benchmark Deep Dive, In Plain English

• SWE-Bench Pro 80.3% - Pro uses held-out, commercial-grade repositories with multi-file changes. Scoring 80.3 there means Fable 5 does not just resolve well-known issues; it navigates unfamiliar, messy, private codebases - which is what your codebase looks like.
• GDPVal-AA 1932 Elo - GDPVal measures professional knowledge work: documents, spreadsheets, analysis products an experienced specialist would produce. A 1932 Elo places Fable 5 first for work that is not coding at all - the model to consider when “junior analyst” is the job description.
• OSWorld 86.0 - Desktop-automation tasks executed against a real operating system. Combined with ExploitBench 78.0, Fable 5 is also the strongest pick for security operations automation we track.

When to Choose Fable 5

Choose Fable 5 when a failed or sloppy attempt costs more than the tokens: production refactors, security-sensitive changes, client deliverables, and anything feeding an autonomous pipeline with limited human review. It is the quality ceiling of the leaderboard across the widest spread of benchmarks. If your workload is high-volume and routine, the same 15K-token request costs 3.7 times less on DeepSeek-V4-Pro-0813 with only 3.5 points less SWE-bench - routing, not loyalty, is the correct answer.

Pricing Reality Check

$14.44 per million tokens in and out makes Fable 5 the most expensive model in our top tier. The same 10,000-request workload we cost other models with (15K tokens average) runs about $2,170. The way to justify that is per solved task, not per token: at 84.5 SWE-bench, failed-attempt rates are low enough that hard tasks get done in fewer round trips. Use it where it wins; see best LLM for coding for the exact crossover math.

Real Limitations

The price is real, and so is the ceiling it creates for high-volume automation - there is no budget tier in Anthropic’s lineup that keeps this capability level. Benchmark breadth also cuts both ways: for pure math, GPT-5.6 Sol posts a higher MATH-500 score, and for raw cost efficiency the open-weights field is not close. Finally, 110 tokens per second is fast for its class but not the fastest in our sourced dataset - latency-critical chat products should compare Gemini 3.7 Flash.

Where to Go Next

• Head to head: Claude Fable 5 vs GPT-5.6 Sol
• How its predecessor set the stage: Claude 3.5 Sonnet benchmark history
• Full pricing context: AI Model Leaderboard