Three days after a model was reported watching its own training, there is a leaderboard for models training models. PostTrainBench hands each coding agent four small base models, a single H100 and ten hours to post-train them, and in this first cut OpenAI’s GPT-5.1-Codex-Max, the same Codex line that had just been reported supervising its own training, led the field (leaderboard). The benchmark becomes the corpus’s ruler for the loop: ten days later Altman confirms self-improving systems in production, v1.0 in March 2026 asks outright whether agents can automate their own post-training, and by July OpenAI reports Sol post-training Luna at 50.3% on it.
Development
PostTrainBench ranks models at post-training other models
The author's Darwinian training loop acquires a scoreboard: once agents are ranked on how well they train other models, the recursion becomes a measured competition rather than an anecdote.
Claim
The new PostTrainBench showed GPT 5.1 Codex Max leading at the recursive task of post-training other models.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to December 18, 2025 (theinnermostloop.substack.com)
- PostTrainBench: Measuring how well AI agents can post-train language models (posttrainbench.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
Reported in
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.