OpenAI's GPT-5.6 Sol autonomously fine-tunes smaller AI
Curated by the Inblix editorial team
OpenAI’s latest model, GPT-5.6 Sol, just flexed an impressive new trick: it autonomously post-trained a smaller model called Luna with minimal human input. A researcher gave Sol a “fairly underspecified prompt” through the Codex platform, and the model independently figured out the training configurations, selected the right GPUs, launched the training script, and verified everything was running smoothly. Previously, this kind of work would have required a team of senior researchers and extra weeks according to OpenAI employee Jason Liu. The milestone is backed by a new internal benchmark measuring Recursive Self-Improvement (RSI)—the ability of an AI to evolve itself. GPT-5.6 Sol scored 16.2 points higher than its predecessor GPT-5.5 on this metric. While full autonomous self-improvement isn’t here yet, this marks a significant step toward AI systems that can accelerate AI research on their own. Why it matters: As AI labs race to use AI to build better AI, this demonstration suggests we’re getting closer to a feedback loop where models can meaningfully accelerate their own development—a key milestone that could reshape the entire pace of progress.
💡 Key Takeaways
- GPT-5.6 Sol autonomously post-trained a smaller model called Luna using only a brief, underspecified prompt
- The model handled tasks like identifying training configurations, selecting GPUs, and executing scripts that would normally take a team of researchers two weeks
- OpenAI created a Recursive Self-Improvement benchmark where Sol scored 16.2 points higher than its predecessor GPT-5.5
Keep reading: See related articles below for more coverage on this topic.
Get smarter about AI
The sharpest AI news, curated daily. Delivered free to your inbox.