- OpenAI says GPT-5.6 Sol independently post-trained its smaller Luna model, finding training configurations, selecting GPUs, and running the job from a brief prompt.
- Sol scored 16.2 points higher than GPT-5.5 on OpenAI’s internal Recursive Self-Improvement (RSI) benchmark.
- An OpenAI employee clarified Sol adapted an existing configuration rather than inventing a recipe from scratch — still estimated to save two researchers about two weeks.
- Recursive self-improvement is central to AI safety debates; Anthropic said in June that full RSI hasn’t been achieved but “could come sooner than most institutions are prepared for.”
What Happened
OpenAI says its flagship model GPT-5.6 Sol independently post-trained the smaller Luna model after Luna’s initial pre-training, according to a July 10, 2026 report from The Decoder. A researcher gave Sol a “fairly under-specified prompt” through the Codex platform, instructing it to find training configurations, pick GPUs, launch the training script, and verify the run.
Why It Matters
AI labs increasingly want to use AI to accelerate their own development, and OpenAI presents Sol as its strongest step yet. “Previously this is something that a team of senior researchers may have worked on at OpenAI, and now it really feels like the automated researcher is pretty close,” OpenAI researcher Kathy Shi said during the presentation. The claim carries safety weight because recursive self-improvement — a system that makes itself better in a compounding loop — has long been a focus of AI risk research.
Technical Details
OpenAI built an internal evaluation suite based on real AI-research tasks — debugging research systems, optimizing kernels and training recipes, running experiments, and improving another model. GPT-5.6 Sol scored 16.2 points higher than GPT-5.5 on the aggregated RSI index, topping OpenAI’s model hierarchy ahead of the Terra and Luna variants and the older GPT-5.5 and GPT-5.4. OpenAI also says average daily token output per active researcher more than doubled its previous GPT-5.5 peak during internal testing, with the share of compute for internal coding inference growing 100x and agent-based token usage roughly 22x — metrics the company concedes don’t directly measure research progress.
Who’s Affected
The claim matters to AI researchers and to the safety community tracking automation of AI R&D. In a follow-up, OpenAI employee Jason Liu added context: Sol didn’t devise a full recipe from scratch, since most of the configuration already existed from Sol’s own post-training; the task was adapting that setup for Luna and running the job — work that would otherwise have “taken two staff researchers maybe an extra two weeks, so this is still a huge deal.” That framing tempers the “automated researcher” language.
What’s Next
Full recursive self-improvement — an AI designing its own successor without human help — has not been achieved. Rival Anthropic said in early June that Claude can now handle incremental work between major paradigm shifts, with humans responsible for only a single-digit percentage of directional decisions, and warned full RSI “could come sooner than most institutions are prepared for.” The measurable next step is whether Sol’s autonomy extends beyond adapting an existing configuration to genuinely novel training design.
