OpenScience scores 75.7% on Terminal-Bench Science, outperforming Codex

ME AI News: Synthetic Sciences, a company in the Y Combinator 2026 Winter Batch (YC W26), has officially launched OpenScience, an open-source research agent. It can search for papers, process data, write code, and interact with scientific databases. The newly added Autoresearch feature enables the AI to autonomously run consecutive experiments. For example, when training a model, researchers only need to say, “Lower the validation loss.” OpenScience will first establish a baseline, then propose modifications on its own, execute experiments iteratively, retain improvements, revert setbacks, and decide the next steps based on prior results. Users can also set limits on the number of runs, total duration, and stopping conditions—or specify constraints like “Only modify the optimizer from now on.” OpenScience automates the iterative research process that previously required constant human oversight: proposing hypotheses, executing experiments, comparing metrics, discarding failed approaches, and advancing to the next round. Experiments can run locally or be delegated to remote GPUs, SSH servers, or Slurm/PBS clusters. Every round’s code, results, and decisions are logged for easy auditability. It supports direct integration with ChatGPT/Codex subscriptions, allows connection to personal API keys or local models, and offers access to the official pay-as-you-go Ace service. In internal tests, OpenScience completed 53 out of 70 Terminal-Bench Science tasks, achieving a score of 75.7%. By comparison, the publicly reported score for Codex + GPT-6 Astra is 68.1%. (Source: BlockBeats)