Synthetic Sciences releases open-source research Agent OpenScience, saying it can iterate experiments automatically

Article

OpenScience Self-Test 75.7%Beyond Codex: A research Agent starts running experiments in a self-directed loop

Beating AI news flash, Y Combinator 2026 winter batch (YC W26) company Synthetic Sciences officially released the open-source research Agent OpenScience. It can search papers, process data, write code, and call research databases. The newly added Autoresearch also allows AI to run experiments continuously on its own.

For example, when training a model, researchers only need to tell it to “reduce the validation-set loss.” OpenScience first runs a baseline, then proposes modifications and executes experiments round by round. It keeps improvements, rolls back regressions, and decides what to try next based on the preceding results. Users can also set limits in advance on the number of runs, total duration, and stopping conditions, or specify that “only the optimizer may be changed next.”

OpenScience automates the experimental iteration that would originally require researchers to keep a close eye on it: propose an approach, run the experiment, compare metrics, discard failed approaches, and continue to the next round. Experiments can run locally or be handed off to remote GPU、SSH servers and Slurm/PBS clusters. The code, results, and decisions from every experiment round are recorded for later review.

It also supports signing in directly with ChatGPT/Codex subscriptions, and can connect to your own API Key, local models, or the official pay-as-you-go Ace.

In the official self-test, OpenScience completed 70 out of Terminal-Bench Science tasks, scoring 53. By comparison, the publicly reported score for Codex +75.7%Astra was GPT-6.

Original link 68.1%https://m.theblockbeats.info/flash/369368

Key points

Synthetic Sciences officially released the open-source research Agent OpenScience.
OpenScience can search papers, process data, write code, and call research databases.
Autoresearch Supports AI continuously proposing approaches, running experiments, and comparing metrics, while keeping improved results and rolling back degraded results.
Users can limit the number of runs, total duration, and stopping conditions, or restrict changes to the optimizer only.
Experiments can run locally, on remote GPU、SSH servers, and on Slurm/PBS clusters.
OpenScience records the code, results, and decisions from every experiment round.
The official self-test says OpenScience completed 70 tasks, of which Terminal-Bench Science 53 were completed, scoring 75.7%。
Evidence says the public score for Codex and GPT-6 Astra was 68.1%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

You can think of it as a research assistant that repeatedly runs experiments on its own: first establish a baseline, then modify the approach, run tests, and compare results; if the results worsen, return to the previous approach.

Why it matters to readers

It links proposing approaches, running experiments, comparing metrics, and eliminating failed approaches into a continuous workflow, which may reduce the work researchers spend repeatedly monitoring experiments. Its actual effectiveness still depends on whether the tests can be reproduced, operating costs, and deployment conditions.

Beginners can understand it as an “automated trial-and-error research programming assistant,” but 75.7%is only an official self-test score and cannot be directly equated with general capability.

Risks and unknowns

75.7%The results come from an official self-test; the evidence does not provide independent verification.
The evidence does not provide the complete test configuration used when comparing OpenScience with Codex.
Remote GPU、SSH The evidence does not specify the stability, cost, and security of server- and cluster-based operations.
The evidence does not specify whether the actual experiments can be reproduced.

Synthetic Sciences releases open-source research Agent OpenScience, saying it can iterate experiments automatically

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center