Anthropic's Claude Fable 5 model tops all six new industry-specific performance benchmarks from Artificial Analysis, covering finance, law, and medicine. However, the cost per task is $3.48, over a hundred times more than DeepSeek V4 Pro's $0.03, despite only a 12-point score difference.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Benchmarking platform Artificial Analysis has released six new industry performance indices. Anthropic's Claude Fable 5 ranks first in all categories, but its cost per task is more than 100 times higher than competitors, sparking debate about value for money.
Artificial Analysis has released six new industry performance indices covering Finance & Accounting, Legal, Healthcare, Strategy & Operations, Engineering, and Economics. These indices are based on task definitions from the U.S. O*NET occupational classification, evaluating models through weighted combinations of domain-specific skills. Anthropic's Claude Fable 5 (with Opus 4.8 fallback enabled) ranks first across all six indices. Claude Opus 4.8 (max) ranks second in five of them, while OpenAI's GPT-5.5 (xhigh) takes second in the remaining two.
On the independent Arena platform, according to the July 7 leaderboard, Claude Fable 5 ranks first in the Text, Code, and Agent arenas—the only lab to lead all three major categories. In the Agent arena, Fable 5 scores 16.58% above the model average, far exceeding GPT-5.5's 8.66% and GLM-5.2's 6.62%.
Among open-weight models, Zhipu AI's GLM-5.2 (max) leads in five industry indices, with an Engineering score of 53, ranking fifth and trailing Claude Sonnet 5 and GPT-5.5 by just 2 points. DeepSeek V4 Pro (max) leads open-weight models in Strategy & Operations with a score of 38.
Rankings vary significantly by domain. Google's Gemini 3.5 Flash, Gemini 3.1 Pro Preview, OpenAI's GPT-5.5, Anthropic's Claude Sonnet 5, and GLM-5.2 swap positions in the second tier.
Top performance comes at a high cost. DeepSeek V4 Flash (max) processes all six index tasks for under $0.04 per task, scoring in the mid-range. GLM-5.2 offers the best performance among open-weight models at $0.26 to $0.58 per task. In contrast, Claude Fable 5 costs up to $3.48 per task on the Strategy & Operations index—over 100 times more than DeepSeek V4 Pro (max) at $0.03, with only a 12-point performance lead.
This sparks a value-for-money debate for enterprises. One strategy is to use a powerful orchestrator model to distribute tasks to cheaper worker models, or to first verify task solvability with a frontier model and then find the cheapest model capable of completing it.
American economist and LLM enthusiast Ethan Mollick believes that Fable 5's performance leap may not be fully captured by current benchmarks, but there is no evidence to support this view. Additionally, GPT-5.6 is scheduled for release the day after the report, potentially altering the competitive landscape.
The benchmark methodology itself has limitations: it is based on static task definitions from occupational classifications, which may not fully reflect dynamic industry needs. Artificial Analysis claims all benchmarks are run independently and the methodology is public, but third-party verification is still needed.
Information comes from third-party benchmark platform Artificial Analysis and independent platform Arena. Data is publicly available and methodology is verifiable, but benchmarks themselves may not fully capture real-world model capabilities.
Claude Fable 5 offers the strongest performance but at an extremely high price. Enterprises must weigh whether a modest performance gain justifies a hundredfold cost increase, or adopt a hybrid model strategy.
Primary report
Primary source