StepFun ships Step-DeepResearch, a single-agent research system

StepFun released Step-DeepResearch, an AI research agent built around a single decision-making loop with what StepFun calls self-correcting logic: adaptive planning and cross-source verification, meaning the system checks its own findings against other sources as it works, rather than routing tasks between multiple coordinating agents. StepFun grounds it in a search index the company says covers more than 20 million papers and over 600 curated indices.

StepFun reports 61.42% on a benchmark called Research Rubrics, which it says matches Gemini's score and beats OpenAI's, and a 67.1% win-or-tie rate against other systems on ADR-Bench. Both figures come from StepFun's own reporting rather than an outside evaluator, and the specific competitor systems in each comparison were not named in the announcement. StepFun published a technical report and code alongside the beta API, allowing outside verification of the architecture claims even though the benchmark numbers themselves remain self-reported.