Anthropic Jumps To Top Of Artificial Analysis Intelligence Index With Fable 5.2, Beats Gemini 4 Carbon By 5 Points

Anthropic's Claude Fable 5.2 scored 65 on the Artificial Analysis Intelligence Index at max effort, five points ahead of Google's Gemini 4 Carbon (60) and six ahead of Claude Opus 5.5 (59), reclaiming the top spot roughly two weeks after Gemini 4 Carbon took it. Artificial Analysis attributed Fable 5.2's gains to scientific reasoning, knowledge and FrontierCode 2.0, with the model setting the highest score on five of the index's nine evaluations, including HLE-Diamond at 80% versus 69% for Gemini 4 Carbon and AA-Omniscience v1.1 at 67 versus 54. At high and xhigh effort Fable 5.2 scored 62, ahead of every other model at any effort setting, while using about 98,000 output tokens per task versus about 173,000 for Opus 5.5 at max effort.

Anthropic has taken back the top spot on the Artificial Analysis Intelligence Index. Anthropic’s newly-released Claude Fable 5.2 scores 65 on the index at max effort, five points ahead of Google’s Gemini 4 Carbon high effort, 60 and six ahead of Claude Opus 5.5 max, 59 . OpenAI’s GPT-6 Astra max sits at 57, and Fable 5.2 is about 15 points above its predecessor, Fable 5.1 max , which scored 49. The result comes barely two weeks after Gemini 4 Carbon took the number one spot https://officechai.com/ai/googles-gemini-4-carbon-tops-artificial-analysis-intelligence-index-becomes-top-ai-model-in-the-world/ with a score of 59, and about seven weeks after Opus 5.5 jumped to the top https://officechai.com/ai/claude-opus-5-5-creates-5-point-lead-over-gpt-6-astra-jumps-to-top-spot-on-artificial-analysis-intelligence-index/ with a five-point lead over GPT-6 Astra. Both of those scores are a point higher now 60 for Carbon, 59 for Opus 5.5 , which appears to reflect the index’s v5.0.1 revision. The lead has changed hands often since September. Artificial Analysis says Fable 5.2’s gains are driven by scientific reasoning, knowledge and FrontierCode 2.0. The three highest-scoring entries on the entire index are all Fable 5.2, at max, xhigh and high effort. Leads five of nine evaluations The Intelligence Index combines 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. Fable 5.2 sets the highest score on five of them. | Evaluation | Fable 5.2 max | Previous best | |---|---|---| | HLE-Diamond | 80% | 69% Gemini 4 Carbon | | AA-Omniscience v1.1 | 67 | 54 Gemini 4 Carbon | | Terminal-Bench-Science 0.1 | 73% | 65% Gemini 4 Carbon | | CritPt v2 Hard Subset | 77% | 72% GPT-6 Astra | | FrontierCode 2.0 | 49% | 46% Claude Opus 5.5 | The HLE-Diamond gain is biggest on knowledge questions, where Fable 5.2 scores 75% against 59% for Carbon. On reasoning questions the gap is narrower, at 86% against 80%. On AA-Omniscience, which measures knowledge reliability and hallucination, Fable 5.2 is 13 points ahead of Carbon, which had itself topped that benchmark by a wide margin https://officechai.com/ai/gemini-4-carbon-tops-updated-aa-omniscience-v1-1-benchmark-with-big-lead-over-competitors/ just over a week earlier. It is also 19 points ahead of Opus 5.5. FrontierCode 2.0 is Cognition’s private evaluation of code that maintainers would actually merge. Fable 5.2 scores 49% there, three points ahead of Opus 5.5 at max effort, and high effort matches max at 49%. Less progress on agentic work Fable 5.2 doesn’t win everywhere. Artificial Analysis says it improves on Fable 5.1 across capabilities, though sometimes only to Opus 5.5 levels. - AA-Briefcase: On this agentic knowledge-work benchmark, Fable 5.2 reaches 1785 Elo. That is within the confidence interval of Opus 5.5 1800 and just behind Claude Sonnet 5.5 1813 . - GDP.pdf: Fable 5.2 reaches 24% all-pass, against 26% for Opus 5.5. - Terminal-Bench 4.0: It scores 60%, level with Opus 5.5 and behind Kimi K3.1 Pro 67% , Gemini 4 Carbon 65% and Sonnet 5.5 64% . - AutomationBench-AA: This is the exception. Fable 5.2 scores 76%, ahead of Opus 5.5 70% but narrowly behind Gemini 4 Carbon 78% . Lower efforts still lead One of the more striking findings is how well Fable 5.2 does without maxing out. At high and xhigh effort it scores 62, ahead of every other model at any effort setting, while using about 98,000 output tokens per task. Opus 5.5 at max effort uses about 173,000 tokens per task and scores 59, so Fable 5.2 at high effort scores higher with roughly 40% fewer output tokens. At low effort, Fable 5.2 scores 54, five points above Fable 5.1 at max, while using about 70% fewer output tokens. Cost per task The benefit at max effort comes at a price. Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, a 20% cut from Fable 5.1. Even so, Artificial Analysis puts its average cost per task at max effort at $26.5. That is about 2.6 times the $10.0 for Opus 5.5 at max, and about twice the $12.99 for Fable 5.1 at max. Artificial Analysis notes that at lower effort settings, Fable 5.2 can outperform Opus 5.5 at a lower cost per task, so the lower settings are where it offers the best value. Other pricing and model details from Artificial Analysis: - Cache pricing: Cache writes cost $10 per million tokens with a five-minute lifetime and $16 with a one-hour lifetime. Cache reads cost $0.20. - Context window: One million tokens, with image and text input. - Effort settings: There are five: low, medium, high, xhigh and max. Artificial Analysis ran the Intelligence Index at all five, with Anthropic’s default fallback enabled. What the chart doesn’t show In Artificial Analysis’s charts, the Gemini 4 Carbon bar is marked as having no publicly available API, a label that also applies to Meta’s Muse Wonder Full. Fable 5.2, by contrast, launched on all platforms https://officechai.com/ai/fable-5-2-benchmarks/ . Among models developers can actually call today, Fable 5.2 leads Opus 5.5 by six points at max effort, and still by three at high effort. Artificial Analysis ran Fable 5.2 with Anthropic’s default fallback enabled, which means requests that trigger its cybersecurity and biology safeguards are handed to a different model. The index therefore reflects Fable’s safeguarded behaviour. Anthropic’s less restricted Mythos 5.2, which shares the same underlying model but is limited to vetted organisations, doesn’t appear in the ranking. The Intelligence Index is one aggregate view, and leadership varies by task. Fable 5.2’s weaker showing on Terminal-Bench 4.0 and AA-Briefcase suggests Anthropic’s top model still doesn’t dominate agentic terminal work or knowledge work, even as it pulls well ahead on science, knowledge and coding quality.