Claude Fable 5.2 Takes the Top Spot: Max Tier Costs 2.6x More, High Tier Is the Better Bet

Anthropic released Claude Fable 5.2. Artificial Analysis placed it in Intelligence Index v5, where the Max tier scored 65 and ranked first. Behind it are Gemini 4 Carbon high at 60 and Claude Opus 5.5 max at 59. A gap of 5 to 6 points is not small.

First, some background. Artificial Analysis is an independent model evaluation organization. Intelligence Index v5 includes 9 evaluations covering science, knowledge, coding, and agentic work. Fable 5.2 leads in 5 of the 9. Its strengths are concentrated: HLE-Diamond 80%, versus the previous best of 69% for Gemini 4 Carbon; AA-Omniscience v1.1 67, versus the previous best of 54; Terminal-Bench-Science 73%, versus the previous best of 65%; CritPt v2 hard subset 77%, versus the previous best of 72% for GPT-6 Astra; FrontierCode 2.0 49%, versus the previous best of 46% for Opus 5.5. FrontierCode 2.0 comes from a private evaluation by Cognition and measures whether code would actually be merged by maintainers. Science, knowledge, and code merge rate—these three areas are where Fable 5.2 pulls ahead.

But several other areas are not as strong. AA-Briefcase is agentic knowledge work; Fable 5.2 gets 1785 Elo, within the confidence interval of Opus 5.5's 1800 and also below Sonnet 5.5's 1813. GDP.pdf full pass rate is 24%, versus 26% for Opus 5.5. Terminal-Bench 4.0 is 60%, on par with Opus 5.5, and behind Kimi K3.1 Pro's 67%, Gemini 4 Carbon's 65%, and Sonnet 5.5's 64%. AutomationBench-AA is an exception at 76%, above Opus 5.5's 70%, but slightly below Gemini 4 Carbon's 78%. So being first on the leaderboard does not mean being first in every scenario. Science and knowledge items raise the ceiling, while agentic knowledge work and terminal coding only reach around Opus 5.5.

Pricing and effort are another key point. Fable 5.2 is priced at $8/$40 per million input/output tokens, 20% lower than Fable 5.1. But the Max tier's Cost per Task is $26.5, about 2.6x that of Opus 5.5 max and 2x that of Fable 5.1 max. Unit price is down, but task cost is not; the Max tier uses more reasoning and output tokens. Effort can be understood as the reasoning investment tier, from low to max in five levels. The higher the tier, the more thinking, time, and tokens are used.

What is really worth looking at is the High tier. High and xhigh both score 62, leading any effort level of other models. The High tier uses 98k output tokens per task, while Opus 5.5 max uses 173k. On cost, High is about $9.08 per task, xhigh about $14.83, and Max about $26.54. xhigh and High score the same but are slower and more expensive. Some commenters said outright that paying for xhigh just means waiting longer for the same score. The low tier is not that bad either: low scores 54, above Fable 5.1 max's 49, at about $3.10 per task. For ordinary knowledge Q&A and lightweight tasks, the low tier offers better value.

Product details: 1M token context window, supports image and text input. Cache write with 5-minute TTL is $10, 1-hour TTL is $16, cache read is $0.20. Effort has five tiers: low, medium, high, xhigh, max. Artificial Analysis ran all five tiers in its evaluation, and Anthropic's default fallback is enabled. The full comparison is on Artificial Analysis's release page: https://artificialanalysis.ai/models/releases/claude-fable-5-2

Some people online say Claude has benchmaxxed again, while others ask whether it's benchmaxxed or a real improvement. The leaderboard can show the ceiling for science and knowledge items, but agentic coding has not pulled away from Opus 5.5, and the Max tier's cost is also right there. If you want to try it, the High tier is a more practical starting point. For scientific reasoning and knowledge-intensive tasks, you can look at the gains from the Max tier; for terminal coding and agentic workflows, first use Opus 5.5 and Sonnet 5.5 as baselines, then decide whether to switch. Between being first on the leaderboard and being good for daily use, there is still cost, latency, and scenario fit in between.