Closed AI systems from Anthropic, OpenAI and Google DeepMind have regained a substantial lead over open-weight competitors, according to the latest Arena AI evaluation data. By September 2026, the strongest closed model had a 29-point advantage on Arena’s Elo-based leaderboard.
Claude Opus 5 Max ranked first with a score of 1505. Moonshot AI’s Kimi K3 Max was identified as the leading open-weight model, but remained nearly 30 points behind.
Open-weight parity has reversed
The latest figures mark a sharp change from January 2025, when the leaderboard showed no Elo gap between the best closed and open-weight models. The difference has since expanded steadily, indicating that open-weight systems have not maintained the temporary parity they achieved early last year.
Separate analysis from Epoch AI points to a similar trend. Open-weight models are now estimated to trail closed systems by about four months on the most demanding tasks, compared with a three-month gap in late 2025.
Arena’s rankings are based on more than 10 million user evaluations each month. Rather than relying solely on synthetic benchmarks, the platform draws on repeated comparisons made by users choosing between model responses. Arena began as a UC Berkeley research project in 2023 and has since developed into a commercial evaluation business reporting an annualized revenue run rate of $100 million, reached within eight months of launching paid evaluation services.
Chinese labs dominate open-weight development
Much of the current open-weight competition is coming from Chinese AI companies. Moonshot AI’s Kimi models, Zhipu’s GLM series, DeepSeek and Alibaba’s Qwen are among the leading publicly available systems. However, the latest leaderboard suggests that Anthropic, OpenAI and Google DeepMind continue to advance their closed models at a faster pace.
Arena AI AI Ability Lead Peter Gostev has focused on making differences in model performance easier to interpret, including the contrast between results from domain specialists and those from general users. That distinction can be significant for companies assessing which systems are suitable for real-world deployment.
