OpenAI grades itself on self-improvement

On July 9 OpenAI opened its GPT-5.6 family to everyone, ending a two-week preview that had been gated to about twenty US-government-approved organizations. That is the launch. The quieter line is what OpenAI now says it measures: an internal suite of 'evaluations measuring progress towards recursive self-improvement,' built from real AI-research chores — debugging research systems, tuning training recipes, and, listed plainly among them, 'improving another model.' OpenAI reports its flagship, Sol, scores 16.2 points above the previous generation on that suite.

One of the tasks in that suite is, in OpenAI's own words, 'improving another model' — and OpenAI says Sol did exactly that, post-training the cheapest sibling it then sold. No third party has confirmed it. — OpenAI

The vivid instance the lab attaches to that last task: OpenAI says Sol autonomously did the post-training of Luna, the cheapest model in the new lineup — the shaping-and-tuning pass that turns a raw model into a shippable product. Read the claim carefully. It is OpenAI's own statement, surfaced in its launch livestream and relayed secondhand; no outside party has confirmed that Sol actually did Luna's post-training, and the specific sentence is hard to pin to a fetchable primary document. What is on paper is the framing — a frontier lab now scoring itself, in public, on how well its models can build the next ones.

This isn't the first machine to post-train a machine: a June research run autonomously post-trained a 30-billion-parameter model and climbed a public leaderboard. What's new is a frontier lab claiming its flagship commercial product did the job on a sibling it then sold. Treat it as a data point, not a milestone — the honest version is 'OpenAI says,' and the same launch undercuts the triumphal frame: Sol trails Anthropic's Claude by roughly fifteen points on a hard software-engineering benchmark, and OpenAI reportedly published an argument that a chunk of that benchmark's tasks are broken. A lab trumpeting self-improvement while it loses the head-to-head it then disputes is the tension worth holding.

For everyone else, the practical change is simpler: GPT-5.6 is now live worldwide across ChatGPT, the API, and OpenAI's coding agent, and the API can now write and run its own JavaScript to chain tool calls, which cut token bills by a third to two-thirds for early customers. The self-improvement claim is the headline the safety community will chew on; the general availability is what actually reaches the millions of people already building on OpenAI.