Clarity on the "Simple Codex" harness in Terminal Bench 2.0?
What is the type of issue?
Documentation is missing
What is the issue?
Is there any public information on what exactly the "Simple Codex" harness is? As of Feb 18, 2026, it's the top of the leaderboard by a significant margin, and since the Codex CLI is open source, it would be nice to know what configuration is being used to achieve these scores, since presumably this should be correlated to getting the most out of Codex (the model + harness) as a developer / end-user.
Is it just the default Codex CLI harness (headless) for GPT-5.3 Codex, but with certain interactive tools (ask user question, etc) dropped?
X-posted to the T-bench repo in case that's a better spot for this discussion: https://github.com/laude-institute/terminal-bench/issues/1417
<img width="1296" height="755" alt="Image" src="https://github.com/user-attachments/assets/3664ac38-96b4-4f4d-aa84-a8b420f9ebd2" />
4 Comments
To my knowledge, this report isn't something we publish or have an affiliation with. You'd need to ask the people who publish this leaderboard.
@etraut-openai do you mean that the Codex CLI research/engineering team is not affiliated with the team that wrote the release blog on GPT-5.3 Codex? Just trying to understand what the relationship between the OSS and leaderboard score is (sounds like it's unrelated teams?)
<img width="669" height="215" alt="Image" src="https://github.com/user-attachments/assets/a2f7e03a-510f-46ac-b23a-e66a661af6fc" />
https://openai.com/index/introducing-gpt-5-3-codex/
(thanks for the quick reply btw)
Ah, I misinterpreted your question. I don't know the answer here. I'm going to convert this to a discussion since it's not a bug or feature request.