On November 17, 2026, OpenAI classified GPT-6.2 Astra as High for AI Self-Improvement and GPT-6.2 Sol below High. The category concerns a model’s ability to help accelerate AI research; OpenAI describes High as an early milestone for preparing safeguards ahead of possible fully automated, recursive self-improvement. OpenAI’s system card sets out the assessment.
For context on the earlier model’s focus on computer use, see NeoTeo’s coverage of GPT-6 Astra.
What OpenAI’s High classification means
The High label concerns AI research acceleration, including research that could improve models’ capabilities. OpenAI frames the threshold as preparation for safeguards in case models may eventually carry out recursive self-improvement fully automatically.
What the three second-stage tests measure
OpenAI says five research-engineering evaluations informed its classification. Two initial-stage tests are followed by three secondary-stage tests. The secondary-stage results use different measures, so each score needs to be read alongside its test and cutoff.
| Evaluation | What the result measures | GPT-6.2 Astra result | Indicative cutoff |
| NanoGPT | Mean share of baseline training time saved | 42.87% | 80% |
| PostTrainBench Lite | Mean share of the performance gap closed across 12 tasks in five-hour runs | 74.79% | 90% |
| MLE-Bench Revised | Mean reference percentile across 72 competitions | 96.06 | 90th percentile |
The NanoGPT and PostTrainBench Lite results fell below their indicative cutoffs; MLE-Bench Revised exceeded its cutoff. Its 96.06 is an average percentile against the benchmark’s reference solutions across 72 competitions.
Read each score by its own measure: training time saved, performance gap closed and reference percentile describe different results.
How the results relate to the overall rating
OpenAI calls the secondary-stage cutoffs indicative and says its High determination drew on all five evaluations. The label is the framework’s overall assessment; the three published secondary-stage results show Astra’s performance on specific research-engineering tasks.