GPT-5.6 Sol Trained Luna From an Underspecified Prompt — +16.2 Points on OpenAI's RSI Benchmark
OpenAI’s GPT-5.6 Sol independently ran the post-training pipeline for Luna, the smallest model in the 5.6 family, after a researcher supplied a single loosely specified prompt through Codex.
The prompt instructed Sol to find suitable training configurations, select GPUs, launch the training script, and verify the run was proceeding correctly. Sol executed all four steps without further direction. Luna’s initial pre-training had already completed; Sol’s job was the specialisation pass that shapes a model’s skills and behaviours.
OpenAI researcher Kathy Shi described the result: “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.”
Employee Jason Liu added context. Sol did not invent a training recipe from scratch — it adapted the configuration that already existed from Sol’s own post-training. The novel part was applying that setup to a different model architecture and executing the run autonomously. Liu put the counterfactual cost at “two staff researchers maybe an extra two weeks.”
The RSI Benchmark
To quantify this class of ability, OpenAI built an internal evaluation suite based on real AI research tasks: debugging research systems, optimising kernels and training recipes, running ML experiments, and improving another model. The suite is named for Recursive Self-Improvement — the theoretical capability where each round of gains enables the next.
On that aggregate index, GPT-5.6 Sol scores 16.2 points above GPT-5.5. The full hierarchy, per OpenAI: Sol leads, followed by Terra, then Luna, then GPT-5.5, then GPT-5.4.
No independent reproduction of the RSI benchmark exists. The numbers are OpenAI-reported.
What This Is Not
Anthropic published a note in early June distinguishing partial from full recursive self-improvement. Full RSI means a system designing its own successor without human involvement; Anthropic said that milestone “could come sooner than most institutions are prepared for” but had not yet been reached. Sol’s Luna post-training is not full RSI. A researcher initiated the task. The training configuration was pre-existing. The model did not set its own objective or evaluate its own output in a feedback loop.
What Sol demonstrated is narrower: a frontier model executing a structured ML research workflow end-to-end — configuration selection, infrastructure orchestration, monitoring — from a prompt that would have produced ambiguity for a junior engineer. That is a meaningful capability step. It is not the feedback loop that RSI researchers track.
The Implication
The story OpenAI is building with 5.6 is one of research automation. Sol at the top of the RSI hierarchy, post-training Luna, scoring 16.2 points above its predecessor. The AI lab that fields the strongest self-improvement toolchain gains a compounding advantage: faster iteration, lower headcount per capability gain, and the ability to direct those researchers toward higher-leverage work.
The two-week task running in hours is the current state of that advantage. The next step — Sol improving on Sol’s own post-training — has not been demonstrated publicly.