Artificial Analysis unveils the Intelligence Index v5 with new coding and science benchmarks; 60% of the Index is based on private test sets or held-out answers

The most capable AI models now require more computational resources and generate substantially more output per task.

Artificial Analysis has launched the Intelligence Index v5, a new benchmark for AI model capabilities. This version significantly updates evaluation methods, with 60% of the Index now based on private test sets or held-out answers, increasing the difficulty and specificity of assessments. The Index introduces two new benchmarks: FrontierCode 2.0, focusing on real-world coding tasks evaluated by repository maintainers, and Terminal-Bench-Science, assessing AI agents in scientific research workflows within a terminal environment. Several existing benchmarks, including HLE-Diamond and AA-Omniscience v1.1, have been upgraded for greater accuracy and relevance. Key results show Claude Opus 5.5 leading with a score of 58, closely followed by OpenAI's GPT-6 Astra and GPT-6.1 Sol. The gap between leading proprietary models and open-weight models has widened significantly, with Kimi K3.1 Pro being the top open-weight model but trailing by 12 points. The cost per task also shows a wide variance, with GPT-6.1 Terra being the cheapest and Claude Fable 5.1 the most expensive.

The v5 update marks a critical shift towards private and held-out datasets, which could make it harder for external researchers to replicate findings or identify specific model weaknesses. This move by Artificial Analysis, coupled with the inclusion of more agentic and terminal-based tasks, suggests a growing emphasis on evaluating AI's practical, real-world deployment capabilities rather than just theoretical performance. The widening gap between proprietary and open-weight models, especially in specialized domains like science, indicates that significant investment is still required to match the frontier for many use cases.

The substantial increase in output tokens per task, particularly for agentic evaluations, highlights the escalating computational demands and potential costs associated with running advanced AI models. This trend suggests that efficiency and cost-effectiveness will become even more crucial factors for developers and businesses integrating AI, potentially driving innovation in model optimization and hardware acceleration. The significant cost variation among top models also indicates a market segment where optimizing for cost-performance will be a key differentiator.