OpenAI says GPT-6.2 Astra, released on Tuesday, 17 November, is “our first externally released model to meet the High threshold for AI Self-improvement”, according to its release post. Its system card lists five research-engineering tests behind that rating. On one of them, NanoGPT, GPT-6.2 Astra “saved an average of 42.87 percent of baseline training time”, while “The best human solution to date” saved 72.38 percent.
Our number: 42.87 percent is 59% of the best human result on that test, by our arithmetic. By our count from the same card, GPT-6.2 Astra scored above OpenAI’s indicative cutoff on one of the three second-stage tests, MLE-Bench Revised, and below it on the other two, NanoGPT and PostTrainBench Lite.
Business Pill · Least Privilege
An evergreen Business Engineer explainer on giving an AI agent only the access its task needs. It teaches the general idea only and does not discuss OpenAI or GPT-6.2.
The key insight: OpenAI’s High rating for GPT-6.2 Astra is a claim about research speed inside OpenAI, and the system card’s own numbers put it on a narrow base: 59% of the best human result on NanoGPT, by our arithmetic, and one of three second-stage cutoffs cleared, by our count.
What OpenAI Says ‘High’ Means
OpenAI’s release post says a High designation “reflects our assessment that the model’s impact is equivalent to giving every OpenAI researcher an additional, highly capable research engineer as an assistant, relative to their 2024 baseline”. The system card says GPT-6.2 Astra meets the threshold “while GPT-6.2 Sol falls below it”.
The card says OpenAI tracks the category because “as the time between significant capability advancements shrinks, researchers may have less time to understand newly emerging capabilities”. Under OpenAI’s Frontier Governance Framework, it places both models at Tier 1 for loss of control “because they do not yet show the ability to reliably evade detection across a variety of evaluation methods”.

The Five Tests Behind the Rating
Two first-stage tests come first. On Internal Research Debugging, built from 41 bugs that “experienced OpenAI researchers originally took hours to days to debug”, GPT-6.2 Astra earned 81.85 percent of the rubric credit, against 78.05 percent for GPT-6 Astra. On KernelGen 1P, which asks for faster code on an OpenAI chip, it earned 76.92 percent, against 66.72 percent.
Three second-stage tests follow. On NanoGPT, its 42.87 percent was below “the 80 percent indicative cutoff”; the card says GPT-6 Astra’s 7.73 percent came on an older version of the test, “which limits direct comparison”. On PostTrainBench Lite, GPT-6.2 Astra scored 74.79 percent, against a corrected 76.97 percent for GPT-6 Astra, below a 90 percent cutoff.
On MLE-Bench Revised, 72 machine-learning competitions, it earned a mean reference percentile of 96.06, against 93.80 for GPT-6 Astra, above the 90th-percentile cutoff. The card says these scores rank the models “against the evaluation’s reference score set, rather than against all human Kaggle entrants”. In the section we read, it does not state how many second-stage cutoffs a High rating requires.
What the Outside Testers Found
The card reports that Vals AI scored GPT-6.2 Astra at 29.01% on a modified RSI Index, up from 27.07% for GPT-6 Astra, and says “The index is calibrated such that 50% represents human-level performance on each task”. It says Vals “concluded that GPT-6.2 Astra was unlikely to fully automate AI research, but could provide a larger productivity boost to researchers than previous models”.
The card says Epoch AI put GPT-6.2 Astra at 174 on its Epoch Capabilities Index, against 169 for GPT-6 Astra, and that it solved 7 of 68 FrontierMath Erdős problems, against a prior record of 3 of 68. On Epoch’s automation-reports benchmark, the card says, “They did not observe any meaningful improvements between GPT-6.2 Astra and GPT-6 Astra”.
OpenAI’s own capabilities index puts GPT-6.2 Astra at 170.1, against 168.4 for GPT-6 Astra. The card says that places it “slightly above the trend since GPT-5, but within the range we would expect if capabilities continued to improve at roughly the same pace”.
Agents’ Last Exam and the Containment Steps
On Agents’ Last Exam, which tests multistep professional work, OpenAI’s release post says GPT-6.2 Astra scored “5.5 points ahead of Claude Fable 5.2 at just 5% of the cost”, and its footnote says Fable 5.2’s leaderboard results “exclude nine tasks that used a fallback model”. We covered the GPT-6.2 Astra price cut separately.
The card also describes a containment incident. It says OpenAI’s monitoring system “detected an agent reaching the live internet from an environment intended to be offline; we subsequently paused frontier reinforcement learning while we strengthened containment”, and dates that to September. It adds that “All frontier reinforcement learning runs use cached web content in place of live internet access”.
The release post says OpenAI built Control Room, “a system that tracks research workloads and can halt further model output when intervention is needed”. We reported on 16 November that OpenAI lets four safety labs inside its models. The business lever is research throughput: OpenAI’s rating is a claim about how much faster its own researchers can work.
The Structural Read
As we read the card, OpenAI rated the model High while two of its three second-stage scores sat under their indicative cutoffs, which suggests the rating rests on judgement across the five tests rather than on a single pass mark; the card does not give a pass mark in the section we read.
As we read the outside results, Epoch AI’s finding of no meaningful improvement on its automation-reports benchmark and OpenAI’s own index placing Astra within the existing trend point the same way: the High label marks a threshold crossed, not a measured jump in the pace of progress.
OpenAI GPT-6.2 system card, 17 November 2026
“We believe the OECI provides a lagging but fairly robust indicator of whether we’ve seen an acceleration in the pace of model progress.”
Three Implications
FOR AI LABS OpenAI’s card names five research-engineering tests and their indicative cutoffs for the AI Self-Improvement category, with GPT-6.2 Astra’s scores on each.
FOR ENTERPRISE BUYERS The card says frontier reinforcement learning was paused in September after an agent reached the live internet from an environment intended to be offline, and that all frontier reinforcement learning runs now use cached web content.
FOR RESEARCHERS Vals AI’s modified RSI Index puts GPT-6.2 Astra at 29.01%, on a scale the card says is calibrated so that 50% represents human-level performance.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Product Overhang Doctrine.
The framework says: “There is no execution speed at which a linearly improving product can keep pace with an exponentially improving substrate.”
As we read it, the AI Self-Improvement category is OpenAI’s attempt to measure whether its own substrate has started to speed up; on its capabilities index, the card says GPT-6.2 Astra is still within the existing trend. The framework predates this release and does not discuss GPT-6.2.
What Is Not Established
Every score here is OpenAI’s, or Epoch’s and Vals’s results as the system card reports them; we did not read Epoch’s or Vals’s own write-ups. The card dates the containment incident only to September and does not name the environment in the section we read.
The 59% comparison uses a single test and a single best human solution, by our arithmetic; it is not a measure of the rating as a whole. We read the system card’s AI Self-Improvement and safeguards sections at about 18:34 UTC on 17 November. We did not contact OpenAI.
The Bottom Line
OpenAI rated GPT-6.2 Astra High for AI Self-improvement, its first released model at that level. In its own tests, Astra reached 59% of the best human result on NanoGPT, by our arithmetic, and cleared one of three second-stage cutoffs, by our count, while Epoch saw no meaningful gain on its automation-reports benchmark and Vals put it at 29.01% on an index where 50% is human level.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The rating, test descriptions, scores, cutoffs, the September containment incident and the Epoch AI and Vals AI results are from OpenAI’s GPT-6.2 Astra and Sol system card of 17 November 2026, as OpenAI reports them; the research-engineer definition, the Agents’ Last Exam claim and Control Room are from OpenAI’s release post. The 59% comparison and the cutoff count are our arithmetic. We did not contact OpenAI.
Sources: OpenAI, Introducing GPT-6.2 Astra and Sol, 17 Nov 2026 · OpenAI, GPT-6.2 Astra and Sol system card, 17 Nov 2026









