Atria Dawn Preview narrowly leads BenchLM research ranking

Atria Dawn Preview from Shanghai Artificial Intelligence Laboratory leads BenchLM’s research ranking as of October 29, 2026, scoring 93.9. OpenAI’s GPT-6 Astra follows with 93.8, ahead of GPT-5.6 Sol at 93.4. The ranking includes 129 models with sourced benchmark coverage and combines demanding knowledge tests with evaluations of web research and synthesis.

The leading model is also the highest-ranked open-weight option, making self-hosting a competitive alternative to proprietary APIs. Context capacity distinguishes the closest contenders: Atria Dawn Preview offers 256K, while GPT-6 Astra provides 1.05M. BenchLM recommends the latter for particularly long inputs despite its slightly lower research score.

The research assessment draws on benchmarks including GPQA, Humanity’s Last Exam, BrowseComp, DeepSearch-QA and GAIA. Model selection should reflect the workload: knowledge-heavy evaluations matter for literature reviews, browsing evaluations for source discovery, and FrontierScience for scientific workflows. Actual performance also depends on search tools, retrieval and citation interfaces, which benchmarks capture only partly. Even leading models can fabricate citations or misread sources, so consequential references still require verification.