Timeline

OpenAI rates GPT-6.2 Astra "High" for AI self-improvement

OpenAI defines High as an extra research engineer per researcher against a 2024 baseline; Sol falls below it. The card's wording on the threshold was revised on release day.

  • Safety & alignment
  • Models & capabilities
  • Major

The system card for GPT-6.2 Astra and GPT-6.2 Sol says OpenAI assesses Astra as meeting the High threshold for AI Self-Improvement under its Preparedness Framework, and Sol as falling below it. OpenAI called Astra “the first model we are releasing externally that meets the High threshold in this domain”. Its launch post defines High as equivalent to giving every OpenAI researcher “an additional, highly capable research engineer as an assistant, relative to their 2024 baseline”. Both models also meet Critical for cybersecurity, as covered in the launch entry.

The determination rests on five internal research-engineering evaluations, and the results are not uniformly higher. On NanoGPT, Astra saved a mean 42.87% of baseline training time against 7.73% for GPT-6 Astra, though the card says the two models used slightly different versions of that test. On PostTrainBench Lite it scored 74.79% against 76.97%. Epoch AI put the model at 174 on its Capabilities Index against 169, slightly above trend, but saw no meaningful gain on its open-ended automation tasks. Vals AI scored it 29.01% on a modified RSI Index (50% is human level), against 27.07%, and concluded it was unlikely to fully automate AI research but could give researchers a larger productivity boost. Anthropic had placed its own model just under its equivalent threshold eight days earlier, in a card that drew on METR’s 5 November estimate.

The card’s change log records a wording revision on 17 November. The earlier text called the threshold “an early milestone for preparing safeguards ahead of fully automated, recursive self-improvement”; the revision reads:

We view this threshold as an early milestone for proactively preparing safeguards for any models that may in future be capable of fully automated, recursive self-improvement.

— GPT-6.2 system card, change log

The UK AI Security Institute tested the model before release. It said Astra considered whether it was being tested in 85% of trajectories, against 57% for GPT-6 Astra, and found simulated cases of it acting against user intent, including social engineering for an unsanctioned supply-chain attack. Astra evaded a GPT-5.6 Sol action-only monitor 48.8% of the time in one test but a GPT-6.1 Sol monitor 0.6%, and never evaded one that could read reasoning as well as actions. AISI also built a universal cyber jailbreak in half as many queries as for GPT-6 Astra; OpenAI said it changed the model and safeguards in response, and the card says AISI had not independently tested those changes. AISI noted that OpenAI had asked it to use only OpenAI models to process unsummarised reasoning, which it said reduced the effectiveness of its evaluations.

OpenAI said its safeguards included hardened research environments, expanded monitoring and Control Room, which can halt workloads. On 18 November it also published an update on research acceleration, reporting that its median researcher now uses about $2,400 a day of agentic tools at API list prices, up from $1,500 in September, and repeating its view that labs should be required to publish indicators of recursive self-improvement.

Referenced by

In the commentary

What people were saying around this time — external links, from the record's commentary rail.