An open-source (OSS) AI agent built by a developer recorded a score of 65.2% on TerminalBench. This figure is reported to exceed Google's official score of 47.8% and the 64.3% achieved by the closed-source model Junie CLI [Hacker News].
OSS AI Agent Outscores Gemini-3-flash-preview on TerminalBench
This article is a translation. Read the Japanese original
The developer emphasized that no illicit methods were used. Specifically, they claim that no cheating mechanisms, such as the insertion of {agents/skills}.md files, were employed.
Furthermore, the developer stated that the execution complied with the leaderboard regulations, with no modifications to resources or adjustments to timeouts. They reported using the open-source version published on GitHub as is.
The developer mentioned that they submitted a pull request to the Hugging Face leaderboard, but since there has been no response from the maintainers, they decided to report the results first [Hacker News].
Sources: Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview（HN 393pt・148コメント） (HN Search (backfill), 2026-04-27)