OpenAI reported that GPT-5.6 Sol post-trained its sibling Luna without human direction, the step where a senior team shapes a model into a product, and called an automated researcher pretty close, years ahead of its own schedule (report). Sol scored 50.3% on PostTrainBench with Terra nosing past at 51.5% (scores), the lab’s experiment throughput has doubled this year (throughput), and Noam Brown said he prefers 5.6 to a human intern (The Information). It lands four months after PostTrainBench v1.0 first measured whether agents can post-train themselves and three after OpenAI set a September target for an automated research intern; five days later Weco reports the first evidence of consistent recursive self-improvement. The lab’s own stated test of whether the loop is real — chips flooding to AI-run research — has not yet happened.
Development
A model autonomously post-trains another model
The post-training step that once took a senior team is done by a sibling model, and the lab's own caveat, that compute has not yet flooded to AI-run research, becomes the corpus's test for whether the loop is real.
Claim
OpenAI said Sol autonomously post-trained Luna, once a senior team's job, calling an automated researcher pretty close and years early, with Sol at 50.3% on PostTrainBench, Terra at 51.5% and experiment throughput doubled this year.
Reported metric: 50.3% / 51.5% PostTrainBench
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to July 10, 2026 (theinnermostloop.substack.com)
- OpenAI says Sol autonomously post-trained Luna (x.com)
- GPT-5.6 Sol at 50.3% on PostTrainBench (x.com)
- OpenAI experiment throughput has doubled this year (x.com)
- OpenAI researcher says GPT-5.6 is better at AI research than human interns (www.theinformation.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- PostTrainBench ranks models at post-training other models
- A benchmark measures whether agents can post-train themselves
- An outer-loop agent rewrites its inner researcher seven times unattended
- A chief executive states the recursion as a product roadmap
- A model autonomously rewrites the kernels behind an 80% price cut
- A student distilled from weaker teachers keeps improving anyway
- Self-improving agents rebuild the inference stack serving two open models
- An insider gives the AGI successor a November ship date
- Welcome to July 10, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.