Claude Fable 5.2 Benchmark Review: FrontierCode 2.0, RoboDojo and Terminal-Bench-Science
Anthropic's Claude Fable 5.2 release reports large gains across repository coding, simulated robotics, terminal-based science, visual chart interpretation, and tool-using knowledge work. The headline numbers are strong. The responsible conclusion is narrower: Fable 5.2 is a promising candidate for complex executable workflows, and organizations should now test it on their own tasks.
One correction is important. The initially proposed article title referenced FrontierMath, but Anthropic's Fable 5.2 launch materials prominently report FrontierCode 2.0, not a Fable 5.2 FrontierMath result. This review uses the confirmed benchmark rather than treating the two as interchangeable.
Confirmed release results
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon | |---|---:|---:|---:|---:|---:| | FrontierCode 2.0 | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% | | FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% | | Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% | | RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | Not stated in release table | | Humanity's Last Exam with tools | 74.2% | 65.6% | 67.7% | 57.2% | Not stated in release table | | Chartography | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Fable's FrontierCode value is reported at xhigh effort; Anthropic's release also identifies effort differences for some competing runs. These configuration details limit simplistic ranking.
FrontierCode 2.0: can the model change a real repository?
FrontierCode evaluates demanding software-engineering tasks that require more than producing a plausible snippet. A model must navigate project structure, infer dependencies, edit coherently, use tools, and satisfy executable checks. Fable 5.2's increase from 32.4% to 49.1% is the release's clearest sign of progress in long-horizon coding.
The limitation is transfer. A private monorepo with undocumented services, custom build tools, and weak tests may bear little resemblance to a benchmark environment. Internal mergeability and human review time matter more than the public percentage.
RoboDojo Sim: planning actions in an interactive environment
RoboDojo Sim tests simulated robotics tasks where success depends on perceiving state, choosing actions, and recovering when a plan does not work. Fable 5.2's 64.9% result is more than double Fable 5.1's 27.6% in the published evaluation.
This is evidence about simulated task-solving, not proof that an unsupervised physical robot is safe. Real systems add noisy sensors, timing constraints, mechanical failure, and consequences that a simulator cannot fully reproduce. Physical deployment requires constrained actions, independent safety systems, and human oversight.
Terminal-Bench-Science: science through tools
Terminal-Bench-Science asks models to solve scientific problems in a terminal environment. They may need to inspect files, run code, transform data, and validate results rather than answer from memory. Fable 5.2 reached 74.6%; the access-restricted Mythos variant reached 75.3%.
This result supports Fable's use as a scientific-workflow assistant, but it does not turn generated findings into verified science. Reproducible scripts, provenance, statistical review, and domain-expert interpretation remain necessary.
Chartography and tool-assisted knowledge work
The 73.4% Chartography score indicates improvement in interpreting charts and visually presented data. A 74.2% result on Humanity's Last Exam with tools suggests strong retrieval and tool orchestration on difficult knowledge questions. Both are relevant to research agents that combine text, images, and external operations.
Yet aggregate accuracy can hide dangerous error clusters. Evaluate the chart types, units, languages, and domain conventions your product actually receives.
Five rules for reading AI benchmark tables
- Compare the harness, tools, effort, and sampling—not just model names.
- Separate vendor-reported results from independent replications.
- Look for confidence intervals and task-level distributions where available.
- Measure cost and latency alongside pass rate.
- Never translate a benchmark percentage directly into a production SLA.
Turning benchmarks into a purchasing decision
Create a small “hard set” from real work: failed coding tickets, difficult analyses, noisy charts, and tool flows that require recovery. Blind-review the outputs and score objective completion, factual support, security, latency, and correction minutes. Compare Fable at fixed effort with Opus 5.5 and the current incumbent. The winner is the system with the best risk-adjusted cost per accepted result.
Frequently asked questions
Did Anthropic publish a FrontierMath score for Fable 5.2 in the launch article?
The launch material used for this review reports FrontierCode 2.0. We therefore corrected the requested title instead of asserting an unverified FrontierMath result.
Why does effort level matter?
Higher effort can improve reasoning while increasing token use and latency. Comparing an xhigh run with a medium run is not a pure model comparison.
Are vendor benchmarks useless?
No. They identify promising capabilities and help prioritize tests. They become misleading only when treated as universal production results.
