Anthropic's Claude Fable 5 dominates new industry benchmarks at a steep premium

Benchmarking platform Artificial Analysis has released six new industry-specific performance indices for AI models. Anthropic's Claude Fable 5 leads every category, but cheaper alternatives handle tasks at a fraction of the cost.

How well do current AI models perform on industry-specific tasks in fields like finance, law, or medicine? Artificial Analysis has introduced six new Capability Indices that compare AI models across Finance & Accounting, Legal, Healthcare & Medical, Strategy & Ops, Engineering, and Economics. The new indices expand on the platform's existing Agentic and Coding indices.

According to Artificial Analysis, the methodology is based on occupational classifications from the US O*NET system. Domain-specific skills are derived from the job tasks defined there, covering things like financial modeling, legal research and contract analysis, or clinical decision support. For each domain, the benchmark suite is assembled fresh and weighted by how often a given skill shows up in that industry. Artificial Analysis says all benchmarks are run independently.

Claude Fable 5 leads all eight indices

Anthropic's Claude Fable 5 (with Opus 4.8 fallback) takes first place across all eight indices. Claude Opus 4.8 (max) comes in second in six of eight categories, according to Artificial Analysis, while OpenAI's GPT-5.5 (xhigh) grabs second in the remaining two. GPT-5.6, set to launch tomorrow, could close the gap between Anthropic's current models and GPT-5.5.

Below the top tier, rankings shift considerably by domain. Google's Gemini 3.5 Flash, Gemini 3.1 Pro Preview, OpenAI's GPT-5.5 (xhigh), Anthropic's Claude Sonnet 5 (max), and the Chinese GLM-5.2 (max) trade places depending on the task.

Among open-weights models, GLM-5.2 (max) leads in five of the six industry indices. With 53 points, GLM-5.2 (max) reaches fifth place overall in the Engineering benchmark, just two points behind Claude Sonnet 5 (max) and GPT-5.5 (xhigh), which both score 55. In the Strategy & Ops Index, Deepseek V4 Pro (max) takes the open-weights lead with 38 points.

The Artificial Analysis results track with current data from LMArena, an independent platform where millions of real users rank models through blind comparisons. As of the July 7 leaderboard snapshot, Claude Fable 5 holds first place in the Text Arena, Code Arena, and Agent Arena. Anthropic is the only lab leading all three main categories.

In the Agent Arena, Fable 5 scores 16.58 percent above the model average, well ahead of OpenAI's GPT-5.5 xHigh at 8.66 percent and Z.ai's GLM 5.2 at 6.62 percent. OpenAI sits at 10th in the Text Arena, Google at 7th. DeepSeek lands at 38th.

A small quality edge comes with a much bigger price tag

Top performance comes at a premium, and that's especially true for Anthropic's frontier models. DeepSeek V4 Flash (max) handles tasks across all six indices for less than $0.04 per task, according to the cost analysis, while scoring in the mid-range. GLM-5.2 (max) offers the best open-weights performance at costs between $0.26 and $0.58 per task.

Claude Fable 5, by contrast, costs $3.48 per task in the Strategy & Ops Index. That's over 100 times more than DeepSeek V4 Pro (max) at $0.03, for a lead of just 12 points. A previous benchmark already showed the steep premium for Fable 5 relative to its modest performance gains.

Whether that gap justifies the price is already a live debate among enterprises. One option is to pair models together, using a capable orchestrator to hand off tasks to cheaper worker models that deliver better value per dollar. Frontier models can also serve as a first check on whether a language model can solve a task at all. Once that's confirmed, the next step is finding the cheapest model that still gets the job done.

Some voices, like US economist and LLM enthusiast Ethan Mollick, argue that the performance leap of a model like Fable 5 simply can't be captured by current benchmarks yet. There's no evidence to back that claim so far.

All indices, including full weightings and benchmark components, are available on the Artificial Analysis website, along with the documented methodology.