OpenAI Says GPT-5.6 Sol Trained Its Own Smaller Sibling

Here’s a detail from the GPT-5.6 launch that deserves more attention than it got: OpenAI says its flagship model, GPT-5.6 Sol, independently post-trained Luna — the smallest model in the new family. Post-training is the phase after a model’s basic education, where it gets optimized for specific skills and behaviors. A researcher gave Sol what OpenAI calls a “fairly under-specified prompt” — vague instructions, essentially — and the model figured out the training configuration, picked suitable GPUs, launched the training script, and checked that everything was running correctly. “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.

To measure this kind of ability, OpenAI built an internal benchmark suite based on real AI research tasks — debugging research systems, optimizing training recipes, running machine-learning experiments, improving another model. On the aggregated result, which OpenAI calls the RSI index, Sol scores 16.2 points higher than its predecessor GPT-5.5. The company also says that during internal testing, daily token output per researcher more than doubled, and the share of computing power going to internal AI-assisted coding grew a hundredfold over six months.

What’s behind this? RSI stands for recursive self-improvement — the idea of an AI system making itself (or its successors) better, with each round of gains making the next round easier. It’s a feedback loop, and it’s been central to AI safety debates for years, because a system that genuinely improves itself could, in theory, accelerate very quickly. Worth being precise here: what Sol did is automation of one training step under human supervision, not a machine inventing a better version of itself. Even Anthropic, OpenAI’s more safety-vocal rival, said in June that full recursive self-improvement hasn’t been achieved — while warning it “could come sooner than most institutions are prepared for.” Also worth remembering: the RSI index is OpenAI’s own internal benchmark, graded by the party with the most to gain.

What this means for you: Nothing in your daily tools changes today. But this is one of those signposts worth filing away: the major labs are now openly measuring — and marketing — how well their AIs can do AI research. That’s the mechanism behind the breathless “acceleration” talk, and it’s why model releases keep arriving faster. If the pace of AI news already feels relentless, this story is a hint that the labs are working on making it more so.

Sources

• OpenAI: GPT-5.6
• The Decoder: OpenAI’s GPT-5.6 Sol autonomously post-trained the smaller Luna model