Anthropic's Claude Fable 5.2 Tops 36 Models but Max Reasoning Backfires
Claude Fable 5.2 tops the Vals Index at 71.31% while costing less than Opus 5.5 and GPT-6 Astra, with a surprising twist on reasoning effort.
- Claude Fable 5.2 takes #1 on the Vals Index at 71.31% accuracy
- Wins seven benchmarks including Vibe Code Bench (93.77%) and BioMysteryBench (83.70%)
- Priced at $8/$40 per million tokens, a 20% drop from the prior generation
- High reasoning effort beats max, which triples cost and hurts accuracy
- Weak on Legal Research, Tax Agent Bench, and HLAB-Vals-1.1
- Still refuses tasks at double-digit rates on Code Migration and ProgramBench
Anthropic's latest reasoning model, Claude Fable 5.2, ranked first among 36 models on the Vals Index, which combines coding, finance, and science evaluations. It scored 71.31% and recorded six first-place finishes across component benchmarks, including one tie with Claude Opus 5.5 on Terminal-Bench 4.0. Vals also found that high reasoning effort outscored max, while max raised the per-test cost from $17.94 to $49.62. Anthropic lowered input and output token prices by 20% from the previous generation.
A narrow win at a lower price
In Vals' reported run, Fable 5.2 led Claude Sonnet 5.5 by 0.46 percentage points and Gemini 4 Carbon by 0.86 points.
| Rank | Model | Vals Index score |
|---|---|---|
| 1 | Claude Fable 5.2 | 71.31% |
| 2 | Claude Sonnet 5.5 | 70.85% |
| 3 | Gemini 4 Carbon | 70.45% |
At $17.94 per test, Fable 5.2 sits on Vals' cost-performance frontier: no evaluated configuration both scored higher and cost less. It also cost less than Claude Opus 5.5 and GPT-6 Astra while posting a higher aggregate score. Vals' per-test figure reflects this evaluation suite; production spending will vary with prompt length, output length, and workload.
Anthropic lists the following API pricing and limits:
- Input tokens: $8 per million, down from $10
- Output tokens: $40 per million, down from $50
- Context window: Up to 1 million tokens
- Maximum output: Up to 128,000 tokens
Coding delivers the standout score
Across the component tests, Fable 5.2 placed first outright on ProgramBench and BioMysteryBench, while sharing first place with Claude Opus 5.5 on Terminal-Bench 4.0.
| Benchmark | Score | Placement |
|---|---|---|
| Finance Agent v2.1 | 69.47% | Tied for first with Opus 5.5 |
| Vibe Code Bench | 93.77% | Tied for first with Opus 5.5 |
| EMB v1.1 | 77.38% | Tied for first with Opus 5.5 |
| Terminal-Bench 4.0 | 65.15% | Tied for first with Opus 5.5 |
| ProgramBench | 19.00% | First of 66; Opus 5.5 scored 18.50% |
| BioMysteryBench | 83.70% | First of 31 |
Those six component finishes, plus the overall Index lead, account for seven first-place results. Fable 5.2 also ranked second on the Vals Science Index at 63.74%, behind GPT-6 Astra.
Vals reports the clearest improvement over Fable 5.1 in coding. The 93.77% Vibe Code score, combined with the Terminal-Bench result, makes the model a strong candidate for coding agents and terminal-based software engineering tasks. Teams still need workload-specific tests because benchmark gains may not transfer uniformly to production repositories and toolchains.
Maximum effort costs more and scores less
Vals evaluated four reasoning-effort settings, which control how much reasoning budget the model can spend before answering: low, high, max, and xhigh. High produced the best aggregate result at 71.31% and $17.94 per test. Max fell to approximately seventh place and cost $49.62 per test, about 2.77 times as much.
The max setting's lower score came largely from incomplete runs. Longer reasoning consumed the available time or token budget, and some responses reached internal output limits before producing a final answer. Those failures reduced aggregate performance even when the additional reasoning helped on individual tasks.
- ProgramBench: Max produced the model's best result across the tested effort settings.
- Tax Agent Bench and Legal Research: Max improved results, though the model's absolute scores remained weak.
- Vibe Code Bench: Low matched high and exceeded max at a lower cost.
Tax, legal work, and refusals remain weak spots
Fable 5.2's aggregate lead masks several low placements that matter for domain-specific deployments.
| Benchmark | Score | Placement |
|---|---|---|
| HLAB-Vals-1.1 | 5.00% | Tied for 29th of 37 |
| Tax Agent Bench | 32.09% | 27th of 84 |
| Legal Research Bench | 46.15% | 19th of 93 |
Vals also recorded refusals, meaning the model declined to attempt some tasks. Refusal rates reached double digits in parts of Code Migration and ProgramBench, with lower rates on Terminal-Bench 4.0, Terminal-Bench-Science, Legal Research, and Tax Agent Bench. Agents that must attempt every subtask need explicit retry rules, fallback models, and separate monitoring for refusals and incorrect answers.
Production policy should follow the workload
- Use high effort as the baseline. It delivered the strongest aggregate score and avoided the completion failures seen at max.
- Test low effort for coding. Vibe Code Bench shows that a cheaper setting can preserve accuracy on some workloads.
- Reserve max for validated exceptions. ProgramBench, tax, and legal tasks gained from the larger reasoning budget, but only workload-level tests can show whether the improvement covers the added cost and latency.
- Keep routing and fallbacks. Weak legal and tax placements, along with refusals, make a single-model pipeline risky for tasks that require consistent completion.
Teams evaluating Fable 5.2 should replay representative production tasks at each effort level and compare accuracy, completion rate, latency, refusals, and token cost. The full Vals results provide a useful baseline, while production traces should determine the final configuration.
Comments
