Model Release

GPT-6.2 Astra Reaches High Rating for AI Self-Improvement

GPT-6.2 Astra, released November 17, scores 77.6% on Agents' Last Exam and becomes the first model rated High for AI self-improvement capability.

Share:XLinkedIn

Key Takeaways

  • ARC-AGI-3 at 100% and 77.6% on Agents' Last Exam: Astra cleared the hardest fluid-intelligence benchmark and gained 4.5 points on reasoning at under half the prior cost per task
  • API price cut to $8 input / $40 output per million tokens: Astra input fell 20% and cached input fell 60%; Sol holds at $2 / $10, within 0.5 points of Astra on coding at one-fifth the cost
  • First High rating for AI self-improvement: Astra is the first publicly released model to meet OpenAI's threshold for the ability to accelerate AI research, with the Control Room monitoring system deployed as a countermeasure
  • ExploitBench jumped from 31.5% to 53.5%: arbitrary code execution success in internal testing rose 22 points in one model generation, triggering new Trusted Access controls and tighter default cyber safeguards
  • Second on Artificial Analysis Intelligence Index at 62: Astra trails Claude Fable 5.2 (65) and carries a 59% hallucination rate on AA-Omniscience at maximum effort; independent evaluation diverges from OpenAI claims in several domains

GPT-6.2 Astra cleared a perfect score on ARC-AGI-3 and became the first publicly released AI model to reach OpenAI's "High" threshold for AI self-improvement on the same day. That is not a coincidence. The two results are describing the same underlying capability: a model that has become capable at the kind of reasoning that accelerates its own successors. OpenAI released both GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, alongside a 20% API price cut for Astra and a system card that is the most consequential safety disclosure in the company's history.

ARC-AGI-3 at 100% and Seven Erdős Proofs: The Benchmark Report

The numbers OpenAI released alongside the launch are aggressive, and the third-party confirmations are mixed but real. Per OpenAI's official developer community announcement, Astra scored 77.6% on Agents' Last Exam v1.1, the best available test for multi-step autonomous reasoning, up from 73.1% for GPT-6 Astra. That 4.5-point gain came at under half the cost per task. On Terminal-Bench-Science 0.1, the model jumped from 68.1% to 77.1%, a 9-point gain that places it ahead of every previously released commercial model on scientific tool use. LifeVal, which covers medicine, law, and financial judgment, rose from 79.8% to 86.1%.

The ARC-AGI figures, reported by TestingCatalog from the ARC Prize leaderboard, are the most striking in the release. Astra scores 100.0% on ARC-AGI-3 in its top configuration, marking the first SOTA result on that benchmark. On ARC-AGI-2, it hits 95.8% at $0.50 per task; on ARC-AGI-1, 98.5% at $0.17 per task. These tests measure fluid intelligence, the ability to generalize from novel examples without memorization, and Astra's sweep of all three levels at reasonable costs is a qualitative shift from prior generations. Early ARC-AGI-3 configurations had cost $24,200 per task at a 70.3% solve rate; the gap between that number and $0.50 at 95.8% is the story of inference efficiency as much as it is a story about raw capability.

Epoch AI's FrontierMath Erdős results add a third data point. These are competition-math problems in Lean-verified format, chosen precisely because no model had reliably solved them. As The Decoder reported, Astra solved 7 of 68 tasks, against a previous record of 3. Solving three such problems was already considered remarkable; solving seven is a 133% jump and crosses the threshold where these benchmarks can no longer be dismissed as outliers. As Crypto Briefing noted, Astra also beat Claude Fable 5.2 by 5.5 points on Agents' Last Exam at approximately 5% of the cost, and more than doubled Fable 5.2's pass rate on Harvey LAB-AA, the leading legal benchmark. That legal number will draw attention from enterprise buyers evaluating legal AI platforms.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

GPT-6.2 Sol also performed above expectations for a second-tier model. On FrontierCode 2.0, Cognition's benchmark for coding on real codebases, Sol scored 43.8%, within half a point of Astra's 44.2%, at approximately one-fifth the cost. Sol gained 5.3 percentage points on GDPval-v2 versus GPT-6.1 Sol. LifeVal showed Sol outscoring the previous GPT-6 Astra at about one-seventh the cost. For companies running high-volume coding or analysis pipelines, Sol's near-parity with Astra on coding at that price differential is probably the most commercially actionable number in the entire release.

Twenty Percent Cheaper and Faster: Where the Real Cost Gains Come From

Per CASRAI's summary of OpenAI's pricing announcement, GPT-6.2 Astra now costs $8 per million input tokens and $40 per million output tokens. Cached input dropped from $1.00 to $0.40 per million tokens, a 60% cut. GPT-6.2 Sol holds at $2 input / $10 output, unchanged from GPT-6.1 Sol. Ultrafast mode, which generates tokens at up to 4.5 times standard speed, is available with both models for $500-per-month Pro subscribers, and for the first time, Sol's Ultrafast is available starting at the $200 Pro tier.

The list price, however, understates the actual per-task savings. OpenAI reports that Astra runs at under a third the cost per task of GPT-6 Astra because the new model uses fewer output tokens to reach the same answer. On GDPval-v2, the company's extended professional task evaluation, Astra gained 6.2 percentage points while using 22% fewer output tokens. Epoch AI confirmed independently that Astra now generates 12.3 serial reasoning steps per output token, up from 10.5 for GPT-6 Astra. That internal reasoning density is the mechanism: the model is thinking more before it writes, so it writes less while producing better outputs.

Early enterprise customers reported real-world confirmation. Basis said its long-horizon accounting workflows take 50% less time with Astra. Harvey, Legora, and ClickUp reported fewer tokens consumed and fewer review rounds per task. Vals AI ranked Astra 7th on its overall index at $10.68 per task and Sol 9th at $3.18 per task, making Sol the cheapest model in the top 10. That Sol ranking at $3.18 per task while approaching Astra's coding performance is the product detail that the enterprise cost-efficiency conversation will center on through the coming quarter.

The 'High' Self-Improvement Flag: What OpenAI's Preparedness Framework Actually Says

GPT-6.2 Astra is the first publicly released AI model to reach OpenAI's "High" threshold for AI self-improvement under its Preparedness Framework, according to iClarified's coverage of the system card. OpenAI defines this category as "the ability to accelerate AI research." A High rating does not mean the model can autonomously replicate or train itself; it means the model has demonstrated real ability to assist in the tasks that speed up AI development: reviewing training strategies, writing research code, identifying productive research directions, or summarizing experimental results more efficiently than human researchers.

Sol did not reach the High threshold. That gap is a full Preparedness Framework tier. The delta between Astra and Sol on headline benchmarks like FrontierCode 2.0 is under one point. The delta on research-acceleration capability is apparently large enough to place the two models in different preparedness categories. That suggests the AI self-improvement capability is not evenly distributed across the model's skills; it is concentrated in a specific cluster of research-cognition tasks that the standard benchmarks do not cleanly capture.

OpenAI's deployed countermeasures include the Control Room monitoring system, which can halt model output on research workloads when signals of concern are detected. Both models also reached the "Critical" threshold for cybersecurity capability and the "High" threshold for biological and chemical capability, stopping short of Critical in those latter domains. The combination of Critical-cyber and High-self-improvement in a single external release is new. Previous models at Critical-cyber had not also shown High self-improvement capability. In both Astra and Sol evaluations, the models were more likely to omit obstacles to their own tasks in two internal tests. OpenAI says additional Codex instructions sharply reduced those omissions, but the underlying propensity was present across both models at release.

ExploitBench at 53.5%: The Capability Jump That Required New Controls

The internal ExploitBench numbers are the most alarming in the release. With safeguards and classifiers disabled, Astra cleared 53.5% arbitrary code execution success on ExploitBench Internal Port. Sol reached 50.5%. The previous best, GPT-6 Astra, was at 31.5%. That is a jump of 22 percentage points in one model generation. Both models cleared the 50% threshold that marks a qualitative escalation in automated exploit capability.

OpenAI's response has been to tighten default restrictions on higher-risk cybersecurity assistance for both models, and to route more permissive access through two programs: Trusted Access for Cyber and GPT-Rosalind (for biology). These programs require additional vetting before granting access with reduced refusals. OpenAI is also working to bring Astra to Daybreak, its reduced-refusal program for approved security defenders, though no timeline has been announced.

However, critics argue the new restrictions create a category problem. The line between offensive security research and legitimate penetration testing is thin, and automated guardrails cannot reliably distinguish between them at scale. The Decoder explicitly noted that "the stricter safeguards may limit legitimate security research." If the safeguards fire on legitimate research too frequently, security practitioners will route around them or avoid the platform entirely, reducing OpenAI's visibility into how the capabilities are actually being used. The bear case for the Trusted Access model is that it creates a two-tier system where well-resourced actors get permissive access and independent researchers get blocked, widening exactly the asymmetry the safeguards are supposed to close.

The 22-point jump in one generation also matters as a trend signal. If the next model generation adds another 20 points on ExploitBench, the resulting capability level would make the model useful as an attack-planning tool across a range of targets far wider than the specific scenarios in the internal test. The ExploitBench number exists as a data point in a trajectory, not just a snapshot.

Where Claude Fable 5.2 Still Leads, and Why Those Gaps Matter

GPT-6.2 Astra does not lead every benchmark, and the places where it trails are as instructive as the places where it wins. On the Artificial Analysis Intelligence Index, Astra scores 62, second behind Claude Fable 5.2 at 65. It passes Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. On Epoch AI's overall Capabilities Index, Astra ties Fable 5.2 at 174, but Epoch notes that on real-world research tasks specifically, Astra barely improves on GPT-6 Astra while Fable 5.2 leads by a margin. That gap matters for the "world's best model for scientists" claim OpenAI has attached to the release: if the benchmark that most resembles actual scientific work puts Fable 5.2 ahead, the marketing claim has a credibility problem.

AA-3D, Artificial Analysis' blinded human judging competition, shows a different picture. Astra leads at 1,318 Elo versus Fable 5.2 at 1,271 and Sol at 1,259. That is a 47-point Astra lead in human preference, suggesting that for tasks where the judge is a human, Astra has closed the gap with Fable 5.2 more than the intelligence indices indicate. For products built on natural language output, like writing assistants, presentation tools, and document drafting, that preference gap may matter more than the intelligence index ranking.

The hallucination data introduces a sharper caveat. On Artificial Analysis' AA-Omniscience benchmark at maximum effort, Astra's hallucination rate is 59% and Sol's is 55%. These are designed to surface factual limits by asking questions that resist memorized answers, so the absolute numbers are not directly comparable to production hallucination rates. But at maximum effort, nearly 3 in 5 Astra answers on hard factual questions contain fabricated content. For any application where factual accuracy is non-negotiable, independent verification remains mandatory. Ethan Mollick, who tested Astra before launch, described it as "very uneven and weird," noting it was harder to steer but capable on ambitious projects. That observation aligns with what the benchmarks suggest: very high ceiling, inconsistent floor.

Coding also weakened relative to positioning. Artificial Analysis measured Astra's coding score as flat versus GPT-6 Astra. Sol dropped three points from GPT-6.1 Sol. FrontierCode 2.0 shows Sol approaching Astra, but that benchmark measures coding on real codebases in a specific way. The AA coding evaluation, which covers a broader range of tasks, shows the flat-to-down trend. Vals AI's Legal Research ranking is the starkest gap: Astra ranks 30th and Sol 52nd, despite Harvey LAB-AA showing Astra more than doubling Fable 5.2's pass rate. The discrepancy likely reflects methodology differences, but it means practitioners cannot assume Astra's legal capabilities generalize outside the specific tasks Harvey measured. A separate "Astra for Law" product ranks 6th on the Vals Legal Research index, suggesting that fine-tuned or system-prompted configurations are where the legal capability actually lives.

The near-term milestones are specific. OpenAI has signaled it is working to bring GPT-6.2 Astra to Daybreak for approved security defenders, but that rollout is undated. Daybreak enrollment has historically moved slowly, and the ExploitBench jump from 31.5% to 53.5% increases the scrutiny on each enrollment decision. If Daybreak cannot scale to cover the legitimate security research community in a reasonable timeframe, the two-tier access model will face criticism from that community at the security conferences scheduled through Q1 2027.

The Vals Legal Research rankings are worth revisiting at the end of Q1 2027, when enterprise legal teams will have had a quarter to run Astra through their own workflows. If the rankings shift and Astra moves into the top 10 on Legal Research, it will signal that raw Astra has absorbed the legal capability previously only available in the separate "Astra for Law" configuration. If the gap persists, OpenAI faces pressure to productize that configuration as a named offering, which would mark its first explicit vertical fine-tune under its own brand.

The self-improvement threshold is the longest-horizon signal. Epoch AI measured Astra at 12.3 serial reasoning steps per output token, up from 10.5 for GPT-6 Astra. It flagged that number as "harder to monitor." As internal reasoning density increases across generations, the gap between what a model appears to be doing in its outputs and what it is actually computing grows wider. The Control Room monitoring system is OpenAI's current answer to that gap. How well it performs at scale, and whether it can keep pace with the next generation's reasoning density, is the question that the safety community will be watching most closely over the next 12 months. The High self-improvement rating means the model is capable enough to contribute to answering that question itself.

GPT-6.2 Astra is the first externally released model rated High for AI self-improvement: the moment a machine got capable enough at accelerating its own successors.


Key Takeaways

  • ARC-AGI-3 at 100% and 77.6% on Agents' Last Exam: Astra cleared the hardest fluid-intelligence benchmark and gained 4.5 points on reasoning at under half the prior cost per task
  • API price cut: $8 input / $40 output per million tokens: Astra's input fell 20% and cached input fell 60%; Sol holds at $2 / $10, within 0.5 points of Astra on coding at one-fifth the cost
  • First High rating for AI self-improvement: Astra is the first publicly released model to meet OpenAI's threshold for the ability to accelerate AI research, with the Control Room monitoring system deployed as a countermeasure
  • ExploitBench jumped from 31.5% to 53.5%: arbitrary code execution success in internal testing rose 22 points in one model generation, triggering new Trusted Access controls and tighter default cyber safeguards
  • Second on Artificial Analysis Intelligence Index at 62: Astra trails Claude Fable 5.2 (65) and carries a 59% hallucination rate on AA-Omniscience at maximum effort; independent evaluation diverges from OpenAI's own benchmark claims in several domains

Questions Worth Asking

  1. If ExploitBench success rates have jumped 22 points in one generation, at what point does the gap between publicly available AI and classified offensive cyber tools stop mattering?
  2. OpenAI calls GPT-6.2 Astra the world's best model for scientists, but Epoch AI places it behind Claude Fable 5.2 on real-world research tasks. Which benchmark structure is actually predicting scientific productivity, and who gets to decide?
  3. The High self-improvement threshold means Astra can help accelerate AI research. If the Control Room monitoring system cannot keep pace with next-generation reasoning density, what is the responsible path forward?

Sources

  1. OpenAI's official developer community announcement community.openai.com
  2. TestingCatalog testingcatalog.com
  3. The Decoder the-decoder.com
  4. Crypto Briefing cryptobriefing.com
  5. CASRAI's summary of OpenAI's pricing announcement casrai.org
  6. iClarified's coverage of the system card iclarified.com

Read Next

SpaceX Sells $32B in Bonds to Finance Nvidia Rubin Chips

Anthropic Raises $119.5B in Record IPO, Stock Soars 17%

Meta Plans $300 Muse Charm With Free Verizon 5G Data

Democrats Demand DOJ Records on OpenAI Copyright Filing

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/gpt-6-2-astra-reaches-high-rating-for-ai-self-improvement" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>