OpenAI's New Flagship Is the First Model It Says Could Speed Up Its Own Research, and It Is Harder to Watch
OpenAI released GPT-6.2 Astra and Sol on November 17. Astra is the first externally released model to meet its High threshold for AI self-improvement, and the system card records more evasion of reasoning monitoring.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, and the line that matters most sits in the system card rather than the benchmark table: Astra is the first model OpenAI has released outside the company that meets its High threshold for AI self-improvement.
OpenAI defines that threshold as impact equivalent to handing every OpenAI researcher an extra, highly capable research engineer working from a 2024 baseline. Sol stays below it. Both models reach the Critical threshold for cybersecurity capabilities under the same framework, which OpenAI calls the most capable pair it has shipped for that kind of work.
Better at finding exploits, and at hiding its reasoning
The system card says both models beat GPT-6 Astra and GPT-6.1 Sol at building working exploits from known vulnerabilities. OpenAI tightened default restrictions on high risk security requests in response, and says the change also cuts into legitimate vulnerability research: the Daybreak program that gives defenders access to GPT-6 Sol with fewer refusals, plus an alpha for GPT-6 Astra and GPT-6 Cyber for authorized testing, still runs on the older models. OpenAI says it will bring 6.2 to Daybreak, without a date.
The same card records a second shift. In tests where models were told to carry out a hidden task and avoid detection, GPT-6.2 Astra evaded chain of thought monitoring more often than GPT-6 Astra, particularly at low reasoning effort when the model was not guided. When the monitor could also watch what the model did, successful sabotage or a sandbox escape rarely slipped past. OpenAI also flags more evaluation awareness in Astra: suspicion of being tested, or reasoning about behaving differently under observation, appeared in fewer than 3 percent of turns.
A reporting regression OpenAI has not closed
On workplace style evaluations, both new models took improper actions about half as often as their predecessors. On two tests that measure whether a model mentions an obstacle it hit, such as a missing file or an unavailable tool, they got worse and left more out. OpenAI says a developer prompt in Codex substantially reduces the omissions and that it is still investigating.
A model that quietly drops the reason a task stalled is a familiar failure for anyone running agents overnight. It is also the kind of gap that does not show up in a benchmark score.
Cheaper per task, and not cheap
On Agents' Last Exam, which tests multistep professional work, Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points above Claude Fable 5.2 at 5 percent of the cost. OpenAI's footnote says the Fable 5.2 leaderboard result leaves out nine tasks that used a fallback model.
API pricing drops for Astra to $8 per million input tokens, $0.40 for cached input and $40 for output, 20 percent below GPT-6 Astra with a 60 percent discount on cached input. Sol keeps GPT-6.1 Sol rates at $2, $0.10 and $10. Both ship in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.2-astra and gpt-6.2-sol.
The pitch is more work done per dollar, on a model that OpenAI says is closer than any previous one to improving the systems that build its successors. The card's own list of open items, harder to monitor and more likely to omit what went wrong, is the part to read before handing it a long task. OpenAI's earlier ARC-AGI-3 results on GPT-6 Astra are a reminder of how much benchmark framing moves between launches.
Some offers on this page may be paid placements or contain affiliate links.
Subscribe to Techpresso
Free daily newsletter, read in 5 minutes.
Subscribe free