Artificial Analysis updates Intelligence Index v5 with new frontier-model rankings
Version 5 of the Artificial Analysis Intelligence Index reframes how frontier-model performance is measured, adding private benchmarks and shifting weights toward agentic collaboration and real-world usage.
1 source · cross-referenced
- Artificial Analysis released Intelligence Index v5, updating its methodology to weight knowledge work, coding, science, and general capabilities at 25% each.
- The new index includes benchmarks like AA-Briefcase, AutomationBench-AA, FrontierCode 2.0, Terminal-Bench 4.0, Terminal-Bench-Science, CritPt v2 hard subset, AA-Omniscience v1.1, and HLE-Diamond.
- Version 5 ranks eight closed models above the best open-weight model, Kimi K3.1 Pro, reflecting a shift toward measuring frontier capabilities rather than open-model parity.
- The update introduces private benchmarks, a move the author argues increases trust concerns by reducing transparency and community scrutiny.
- The author notes that leading open labs are expected to close gaps in future model cycles, but cautions against over-reliance on single-number indices.
Artificial Analysis released Intelligence Index v5, an update to its prominent independent intelligence metric for frontier language models. The new version weights four domains equally at 25% each: knowledge work, coding, science, and general capabilities. Within knowledge work, the index combines AA-Briefcase (15%), AutomationBench-AA (5%), and GDP.pdf (5%). Coding is measured by FrontierCode 2.0 (15%) and Terminal-Bench 4.0 (10%). Science uses Terminal-Bench-Science (15%) and the hard subset of CritPt v2 (10%). General capabilities are represented by AA-Omniscience v1.1 (15%) and HLE-Diamond (10%).
The update introduces a new ranking of frontier models, with eight closed models outperforming the best open-weight model, Kimi K3.1 Pro. The author characterizes this shift as a move toward measuring where frontier models are heading—toward general agentic collaboration—rather than where open models are strongest today. This directionally defensible change, however, risks narrowing the benchmark’s relevance to the broader open-model economy, which emphasizes commodity inference, internal data specialization, and controlled agent deployment.
Artificial Analysis also signals a shift toward private benchmarks, where prompts and evaluation details are not publicly disclosed. The author argues this reduces transparency and removes the benefit of community scrutiny that has historically improved benchmark quality. With significant financial stakes tied to these indices, the author favors transparency to enable broader participation in identifying flaws and improving evaluations.
The author expects leading open labs to close gaps in future model cycles, noting a recurring pattern where open models narrow performance differences over time. However, the author cautions that open models may remain narrower in capability breadth, particularly in scientific domains, and that rapid progress across multiple domains requires sophisticated reinforcement learning stacks that not all open labs possess.
The author observes that closed labs like Anthropic and OpenAI have largely deprioritized these public indices in favor of internal evaluations tailored to their product needs. This underscores the growing divergence between public benchmarking and the proprietary evaluation regimes used by leading labs to guide model development.
- Nov 6, 2026 · OpenAI — News
OpenAI releases BioBasicsBench to evaluate AI agents on wet-lab science tasks
Trust79 - Nov 3, 2026 · Hugging Face
Allen Institute releases StateBench to test long-context state tracking in language models
Trust79 - Nov 3, 2026 · Hugging Face
CWE-bench v2 expands public cybersecurity evaluation suite to 100 tasks across 91 CWEs
Trust79