See a question become an answer.
Two minutes: a folder of data and a question, the steps Jeus takes, one published question with every agent's recorded outcome, and the headline results.
The task the benchmark authors chose, and what every agent did with it.
Each benchmark's authors illustrate their paper with one task. We took that task whenever it was in the collection we evaluated, and we show every agent's outcome, whatever it was; the choice was made before any grade was seen. One evaluation per agent, model and question, with Codex CLI, Claude Code and Jeus on the same model, data and limits.
Five open-weight models, three benchmarks, four agents.
Each bar is a five-model mean: GLM-5.3, Qwen3.8-Flash-Next, DeepSeek-V4.1-Flash, MiMo-V2.6-Pro and Kimi-K2.6, one evaluation per agent, model and question. BixBench has 205 questions in 54 data capsules, Verified-50 its 50 expert-checked questions, and BAISBench 193 multiple-choice questions on 41 single-cell studies, scored by its own official rule.
BixBench accuracy, Verified-50 accuracy and the official BAISBench score, in percent. The next agent on each: Codex CLI 49.5, Claude Code 82.8 and Claude Code 65.1.
Peak resident memory of the agent process, as a multiple of Jeus's, over 28 benchmark summaries of Codex CLI, Claude Code and Biomni A1. Model servers are excluded for every agent; Biomni A1 runs its analyses inside its own process.
With Claude Sonnet 5 and GPT-6 Sol, Jeus's recorded model-access charges were lower than the other agent's in 14 of 16 benchmark comparisons, by 24 to 81 percent. Judging is excluded.
We also evaluated 24 clinical prediction tasks, 12 of classification and 12 of survival, with the same agents and models.
Worked examples, three results, and a place to try Jeus.
To try Jeus yourself, sign in with a Google account at the live demo. Ask one question and two agents answer it side by side on GPT-6 Sol and the same files: Jeus, and OpenAI's own agent, Codex CLI. Token usage and time show under every reply, and each agent's total tokens and cost show at the end. You get a private workspace with example data, you can upload files of your own, and each person may use $5 of model usage a day.
Codex CLI, Claude Code and Biomni A1 ran with the same models, the same data and the same limits as Jeus. Results depend on the model behind each agent as much as on the agent; the examples above show both.