Artificial Analysis Rebuilds Its AI Capability Indexes, Crowning Claude Opus 5.5

Independent model evaluator Artificial Analysis has updated its six industry-specific Capability Indexes to v2.0 and aligned their benchmark pool with Intelligence Index v5 . The revision gives more weight to non-hallucination and agentic tool use, measuring whether a model avoids fabricated answers and can operate software inside multistep workflows. Those capabilities affect production risk, automation quality, and model-selection costs.

Occupations set the score

O*NET, the US Department of Labor’s taxonomy of roughly 1,000 occupations, provides the task data behind each index. Artificial Analysis converts that data into domain scores through three steps:

1. Identify the tasks associated with occupations in each industry.
2. Map the capabilities required for those tasks to relevant benchmarks.
3. Weight each benchmark according to how often its capability appears across the task mix.

Each resulting index is a job-weighted average of benchmark results. The six domains are Finance and Accounting, Strategy and Ops, Legal, Healthcare and Medical, Engineering, and Economics.

One benchmark pool, six views

All six indexes now draw from the evaluation pool used by Intelligence Index v5:

• 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)
• Terminal-Bench-Science 0.1

In Intelligence Index v5, evaluations with private test sets or held-out answers account for 60% of the score. Private material reduces benchmark contamination, which occurs when test questions or answers enter training data and inflate measured performance.

Reliability and tools gain weight

Four benchmark changes apply across the indexes:

• HLE-Diamond replaces Humanity’s Last Exam .
• AA-Omniscience advances to version 1.1.
• GDPval-AA leaves the Agentic Knowledge Work category.
• AA-LCR leaves Long-Context Reasoning, with GDP.pdf filling that role in Economics.

Engineering gets a harder science test

The new Engineering science score uses CritPt v2 and Terminal-Bench-Science 0.1 . Artificial Analysis evaluates CritPt v2 on a private 40-question subset drawn from 70 research-level physics questions. It removed 30 questions that seven frontier models answered correctly in all five attempts at launch, concentrating the evaluation on questions that continue to distinguish leading models.

Inference effort changes cost by 10x

At maximum effort, Claude Opus 5.5 ranks first across all six Capability Indexes. GPT-6 Astra at maximum effort places second in Finance and Accounting, Legal, and Engineering. The effort setting controls how much inference-time computation a model uses, creating substantial differences in score, token consumption, latency, and price.

Maximum effort raises the score by about 66% while increasing cost by roughly 10.2 times. Model comparisons therefore need matched effort settings and full task costs, especially when production traffic magnifies small per-request differences.

Unreleased weights complicate the open ranking

Kimi K3.1 Pro leads models categorized as open-weight across all six indexes and ranks between fifth and eighth overall. Its weights remain unreleased, so developers cannot yet reproduce or self-host that deployment. Among models with released weights, MiniMax-M3.1-Flash leads Strategy and Ops at No. 10 overall, while DeepSeek V4.1 Flash 1017 at maximum effort leads Healthcare and Medical at No. 12.

Use v2.0 as a new buying baseline

For model selection, the domain indexes offer a more targeted screening tool than a general composite score. Teams can apply them without treating any leaderboard as a substitute for internal evaluation:

• Select the domain whose O*NET task mix most closely matches the intended workload.
• Compare models at the same effort setting and include the full cost per completed task.
• Test finalists with private prompts, tools, documents, and failure criteria from the production environment.
• Measure end-to-end latency and token consumption alongside benchmark accuracy.

Version 2.0 also creates a new comparison baseline because both the benchmarks and their weights have changed. Artificial Analysis reports that every measured model generated more output tokens per Intelligence Index task under v5 than under v4.3. The median rose by about two-thirds, from 35,000 to 59,000 tokens. Rankings from the v1.x era and earlier token-cost comparisons should therefore be rerun against the new task set before a vendor or deployment configuration is selected.