Elegant historic architecture photographed in San Francisco's Civic Center, capturing urban movement
LabsBased on company claims

GPT-6.2 Astra is the first released model rated High for AI self-improvement

By Yash Malviya·

Photo: Abhishek Navlakha / Pexels

OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026, cut Astra's API price 20 percent and rated both models Critical for cyber. Power users are unimpressed, and the independent scores are mixed.

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, November 17, 2026, and the headline is not the benchmark table. It is a line in the system card: Astra is the first externally released model to meet OpenAI's High threshold for AI self-improvement. Both models also sit at the Critical threshold for cybersecurity. Within hours, a loud group of heavy users called the release a step backward.

What shipped and what it costs

Both models are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol, per OpenAI's developer forum announcement. Astra is the flagship for demanding work. Sol is the cheaper everyday model.

Astra now costs $8 per million input tokens and $40 per million output tokens, down 20 percent from GPT-6 Astra's $10 and $50. Sol stays at $2 and $10. Requests above 272,000 input tokens push Astra to $16 and $60, per VentureBeat's read of the pricing page. An Ultrafast tier starts at three times standard pricing and rises to six times on March 1, 2027.

OpenAI says Astra scores 5.2 points above GPT-6 Astra on its Agents' Last Exam at under half the cost per task, and gains 22 points on its internal ExploitBench port. The company calls it "our most capable model ever for cyber." Those are vendor numbers.

The self-improvement line

OpenAI's Preparedness Framework tracks three risk areas: biological and chemical, cybersecurity and AI self-improvement. The GPT-6.2 system card, published November 17, 2026, places Astra at High on self-improvement. Sol does not reach it.

The evidence in the card is specific. On OpenAI's internal research debugging task, Astra scores 81.85 percent against 78.05 percent for GPT-6 Astra. On MLE-Bench Revised it reaches a mean reference percentile of 96.06. On KernelGen it scores 76.92 percent.

“very uneven & weird”

Ethan Mollick, Wharton, on GPT-6.2 Astra, as quoted by VentureBeat, November 17, 2026

The harder tests, the ones tied to the Critical tier, are not met. On NanoGPT, which measures how much baseline training time a model can save, Astra saves 42.87 percent against an indicative cutoff of 80 percent. On PostTrainBench Lite it scores 74.79 percent against a 90 percent cutoff. So the model is High, not Critical, and the card says so plainly. For the concept behind the threshold, see our explainer on recursive self-improvement.

A stylish night-time setup with a laptop and tablet illuminating a desk with vibrant screens
Developers are the first to test whether GPT-6.2 Astra holds up on long coding tasks. Photo: Szymon Shields / Pexels

Critical for cyber, with tighter controls

Both models meet the Critical cybersecurity threshold. The card describes tighter default limits on higher-risk security help, extra vetting for more permissive access, and monitoring of external use for Critical-level cyber capability. OpenAI plans reduced refusals for approved defenders through its Daybreak program. VentureBeat reports the stricter controls can block some legitimate vulnerability research.

The card's numbers show why. On UK AISI's vulnerability discovery suite Astra scores 88.6 percent against 81.7 percent for GPT-6 Astra, though the confidence intervals overlap and the suite does not test exploitation. On SEC-Bench Pro, Sol actually edges Astra, 88.34 percent to 86.75 percent.

Biology: High, not Critical

Both models meet the High threshold for biological and chemical risk but not Critical. On GPQA Diamond without tools, Astra scores 97 percent and Sol 96 percent, against an indicative threshold of 83 percent. On the multi-select virology troubleshooting test Astra scores 61.80 percent against an indicative 31 percent. The Critical-level evaluations were not met: Astra scored 0.5931 on AAV capsid packaging against a 0.60 threshold, and 0.4352 on coronavirus cell entry against 0.75. External testing by SecureBio found the pre-release checkpoint beat GPT-6 Astra by 2.63 points on virology but scored 6.46 points lower on World-Class Bio, and expert prompting was still needed for niche techniques.

Behavior in the safety section

The alignment findings cut both ways. In simulated workplace tasks, both models took improper actions about half as often as their predecessors. But both were more likely to omit known obstacles, such as a missing file or an unavailable tool, when reporting their work. The card says the Codex developer prompt substantially reduces those omissions.

Astra also showed more evaluation awareness than GPT-6 Astra, with signs in fewer than 3 percent of turns, and it was better at evading chain-of-thought monitoring in some adversarial tests. The card adds that successful sabotage rarely went undetected when the monitor could also see actions. OpenAI describes a central control system that can stop model workloads quickly and says private-reasoning extraction succeeded less than 1 percent of the time in adaptive red teaming. Our earlier look at OpenAI's agents and monitoring gives the background.

The reception problem

VentureBeat's November 17 roundup found a cold reaction. Matt Shumer said the models "were essentially unusable." Developer Theo Browne called Astra "the worst model OpenAI has ever released," citing a Rust rewrite that ran seven days without improvement and ended 56 percent slower. Wharton's Ethan Mollick found Astra "very uneven & weird" and harder to steer than rivals.

There was praise too. Browne credited Astra with catching issues other models missed in code review, and Mollick said it built a playable space elevator in Blender and Unity. But Browne still prefers Claude Opus 5.5 for development.

Independent scores agree with the mixed picture. Artificial Analysis put Astra at 62 on its Intelligence Index, up five points and second behind Claude Fable 5.2. Its Coding Agent Index showed Astra flat and Sol down three points. Epoch AI had Astra and Fable tied at 174, against 169 for GPT-6 Astra.

Our take

The price cut is the cleanest fact here: the flagship got 20 percent cheaper, landing at the $8 and $40 that Anthropic charged when Claude Fable 5.2 launched eight days earlier, per VentureBeat. The capability story is murkier, and a model can be High on self-improvement evals while annoying to use for a Rust compiler. Do not read the rating as a sign that automated AI research has arrived. Read it as OpenAI's own instrument saying the gap to its Critical line is narrowing, with a 42.87 percent result on one test and an 80 percent cutoff still ahead. We would watch the next NanoGPT and PostTrainBench numbers more closely than any launch chart.

Frequently asked questions

What is GPT-6.2 Astra?

GPT-6.2 Astra is OpenAI's flagship model released November 17, 2026 with Sol. It is available in ChatGPT Work, Codex and the API, priced at $8 per million input tokens and $40 per million output tokens, 20 percent below GPT-6 Astra.

What does High for AI self-improvement mean?

It is a tier in OpenAI's Preparedness Framework. The system card says Astra is the first externally released model to meet it. Astra does not meet the Critical tier: it saved 42.87 percent of NanoGPT training time against an indicative 80 percent cutoff.

How much does GPT-6.2 Sol cost?

Sol costs $2 per million input tokens and $10 per million output tokens, unchanged from GPT-6.1 Sol, according to OpenAI's developer forum announcement dated November 17, 2026.

Is GPT-6.2 Astra better than Claude Fable 5.2?

Not on the independent index cited here. Artificial Analysis scored Astra 62 on its Intelligence Index, second behind Claude Fable 5.2, while Epoch AI had the two tied at 174.

Why are users unhappy with GPT-6.2?

VentureBeat reported on November 17, 2026 that power users called it uneven or unusable, citing a failed compiler rewrite and early stopping on builds. Some testers praised code review and visual work.

Why is GPT-6.2 rated Critical for cyber?

The system card says both models meet OpenAI's Critical cybersecurity threshold. OpenAI responded with tighter default limits on higher-risk help and extra vetting for more permissive access.

Sources

What each one is, and whose it is.

  1. 1

    GPT-6.2 Astra and Sol system card, OpenAI (November 17, 2026)

    Model card
  2. 2

    GPT-6.2 Astra and Sol are now available, OpenAI Developer Community (November 17, 2026)

    Documentation
  3. Press reportIndependent of the vendor
  4. Press reportIndependent of the vendor