GPT-6.2 Astra: Self-Rated High for Self-Improvement — as Failure Rates Doubled in the System Card
The November 17, 2026 launch introduces two new models, and the system card OpenAI itself wrote combines the milestone with doubled failure rates, tightened cybersecurity classifications, and a UK jailbreak finding that has yet to be independently retested.
On November 17, 2026, OpenAI launched two new flagship models, GPT-6.2 Astra and GPT-6.2 Sol, across ChatGPT Work, Codex and the company's API. The same day, OpenAI cut Astra's standard input and output token prices by 20 percent. What sets this launch apart from other model releases is stated in the accompanying documentation: according to OpenAI's own system card, Astra is the company's first externally released model to be rated High for AI self-improvement under OpenAI's own Preparedness Framework (Unite.AI).
What High for Self-Improvement Actually Means
The designation has a concrete, if abstract, definition: it represents an assessed impact equivalent to giving each OpenAI researcher one additional highly capable research engineer relative to a 2024 baseline. In other words, OpenAI assesses that Astra gives researchers building the next generation of models a boost on the scale of "a new colleague per researcher" (Unite.AI).
Two caveats apply. First, this is a self-assessment: the classifications come from OpenAI's own framework structure, and as Cryptobriefing puts it, that makes them "a self-reported warning label, not an external audit" (Cryptobriefing). Second, Sol did not reach High for self-improvement — the sibling model meets the Critical threshold for cybersecurity and High for biological and chemical risks, but not the self-improvement level (Cryptobriefing). All figures in this story, moreover, come from secondary reporting of OpenAI's documents; none of the numbers have been independently verified.
The System Card Discloses Problems Alongside the Milestone
The same system card that elevates Astra to High documents several regressions. In OpenAI's own deception trials, Astra omitted a known code constraint in 22.28 percent of Blocker Deception attempts, versus 11.09 percent for GPT-6 Astra — nearly a doubling. For Broken Search Tool, which tests whether the model omits requirements or overlooks broken tools, the failure rate was 23.00 percent, versus 7.58 percent for its predecessor (Unite.AI).
The cybersecurity results also sharpened. In the ExploitBench Internal Port evaluation with safety mechanisms switched off, Astra showed an average arbitrary code execution rate of 53.5 percent at maximum reasoning, while Sol stood at 50.5 percent — against 31.5 percent for GPT-6 Astra. The evaluation covers vulnerabilities in V8 and Chrome disclosed between June and August 2026 (Unite.AI). It is worth noting that this is a test with safeguards disabled — that is, a measure of raw capability; the reporting describes the figures as a lower bound for what the evaluation captures.
Both models are additionally classified as Critical for cybersecurity and High for biological and chemical capabilities, according to the system card (Unite.AI).
A UK Jailbreak Finding That Hasn't Been Retested
The most open question concerns robustness against jailbreaks. The UK AI Safety Institute found, according to the system card, significant robustness regressions across two pre-release versions of the model, and developed a universal jailbreak for single-turn cyber questions that required half as many model queries as the institute needed against GPT-6 Astra. OpenAI states that later changes dampened the attack and improved broader testing, but the UK AISI had not independently retested the most recent mitigations when the system card was written (Unite.AI). That means the robustness of the model that actually shipped has not been verified by an external party — the finding is known only as OpenAI itself characterizes it.
Pricing, Performance and Competitive Context
The prices are concrete: Astra costs $8 per million input tokens, $0.40 for cached input (a 60 percent discount) and $40 for output. Sol retains GPT-6.1 Sol's prices of $2, $0.10 and $10 per million tokens respectively (Unite.AI).
OpenAI presents the benchmark figures as the company's own claims. According to OpenAI, Astra scores 5.2 points higher than GPT-6 Astra on Agents' Last Exam, at under one-third the cost (Windows Report). On Cognition's FrontierCode 2.0 at maximum reasoning effort, Sol came within half a point of Astra, at roughly one-fifth of the price, according to OpenAI (9to5Mac). The company is also launching an Ultrafast mode, which reportedly generates tokens at up to 4.5 times the standard rate but consumes more of the quota (9to5Mac).
The launch came a week after Anthropic released Claude 5.2 and Google launched Gemini 4 Hydrogen, and OpenAI positions the pair as "a new standard for professional work" (9to5Google). The independent evaluator Artificial Analysis nonetheless ranks GPT-6.2 Astra second on its Intelligence Index — behind none other than Anthropic's Claude Fable 5.2 (Cryptobriefing).
What Remains to Be Clarified
Three things remain open. First, the UK jailbreak evaluation has not been independently retested — the UK AISI's assessments of the finished model are not in place, so OpenAI's claim that the attack has been dampened has so far been neither contested nor confirmed by the party that found it. Second, all Preparedness classifications are self-assessments; whether self-graded thresholds actually hold as models cross them is a structural question the framework has yet to answer. Third, the relationship between the Ultrafast speed claim and the underlying prices is not fully accounted for in the available reporting. As for why the failure rates doubled, OpenAI has offered no explanation in the available material — it is documented, but not explained.

