Claude Fable 5.2 posts record 65 on Artificial Analysis Intelligence Index

A new record on the Artificial Analysis index

Claude Fable 5.2 has recorded the highest score to date on the Artificial Analysis Intelligence Index, according to a Hacker News front-page item published on November 9, 2026 that links to the benchmarking service's model page. The model, listed in its max configuration, posted a score of 65. Neither the Hacker News item nor the model page states the margin over the previous leader.

The listing carries a repeated note that the model has no publicly available API. That annotation appears across the intelligence, cost and performance sections of the page, suggesting the score reflects measured capability rather than a product developers can currently call.

What the index measures

The result is presented under version 5.0.1 of the Intelligence Index, which aggregates nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset) and Terminal-Bench-Science 0.1.

According to Artificial Analysis's methodology notes, the mix leans toward agentic and applied work rather than quiz-style knowledge questions. The component evaluations include agentic software engineering in open-source repositories, scientific research workflows carried out inside a terminal, and professional document reasoning under the GDP.pdf test.

Two components are described in notable detail on the page. AA-Briefcase v1.1, developed in-house by Artificial Analysis, is an agentic knowledge-work benchmark reported as an Elo rating that combines rubric pass rates with separate judgements of analytical quality and presentation. AA-Omniscience v1.1 yields a net score between -100 and 100 that rewards correct answers, penalises hallucinations and imposes no penalty for refusing to answer; a score of zero means correct and incorrect answers cancel out.

Cost and speed figures are absent

Artificial Analysis normally pairs its intelligence score with economics and throughput: a weighted average cost per index task built from input, cache-hit, cache-write, reasoning and output token prices, output tokens per task, output speed, latency to first token and context-window size. The service says it reruns the benchmark every six hours, averages speed and latency over the previous five days of endpoint runs, and checks endpoints of deprecated models every 18 hours.

For Claude Fable 5.2, that surrounding data cannot be populated while no public API exists, so the headline number stands without the price-and-speed context the tracker normally provides for reachable models.

Caveats for readers

The score is specific to index version v5.0.1 and its particular weighting of nine benchmarks, so it will not map cleanly onto other leaderboards or earlier index versions. The heavy agentic component also means results may favour models that sustain long, tool-using workflows, which is not the same as strong performance on every real-world task. Finally, the available source material is a model page plus a Hacker News summary, so details such as pricing, throughput and the gap to rival models were not available at the time of writing.

Why it matters

A record on the Artificial Analysis Intelligence Index is one of the more widely watched signals in frontier-model competition, because the benchmark tries to compress reasoning, coding and agentic performance into a single comparable number. A new high-water mark will inevitably shape expectations for whatever competing labs ship next.

The second half of the story is access. A model can top an independent leaderboard and still be unusable if there is no documented endpoint, and that is exactly the situation the listing describes. For developers, the practical question is no longer whether Claude Fable 5.2 is capable, but when it can be reached, at what price and at what speed.

The benchmark composition is also a signal in itself: evaluators are steadily moving away from static question sets toward agentic, multi-step tasks that look more like actual work, and leaderboard leadership is increasingly being defined on those terms.