OpenAI releases GPT-6.2 Astra and Sol: cheaper, faster — and its first model rated High for AI self-improvement
OpenAI shipped GPT-6.2 Astra and Sol on November 17, with higher benchmark scores at lower prices — and Astra is the company’s first model rated High for AI self-improvement.
- The models: GPT-6.2 Astra, the frontier model for the hardest projects, and GPT-6.2 Sol, the efficient model for everyday work. Released November 17, 2026 in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, plus the API as gpt-6.2-astra and gpt-6.2-sol. They are not in regular Chat.
- The scores: on Artificial Analysis’s Intelligence Index, Astra 6.2 scores 62 (up from 57) and Sol 6.2 scores 57 (up from 55). Weighted cost per index task fell to $4.00 from $6.24 for Astra and $1.01 from $1.27 for Sol. On FrontierCode 2.0, Sol comes within half a point of Astra at maximum reasoning effort for about one-fifth of the cost.
- The prices: Astra costs $8 per million input tokens and $40 per million output tokens, 20% below its predecessor, with cached input at $0.40. Sol is unchanged at $2, $10 and $0.10. Ultrafast reaches Pro 200 subscribers (with Sol) for the first time, and OpenAI is cutting Ultrafast pricing by 50% for the next three months.
- The safety ratings: both models are rated Critical for cybersecurity under OpenAI’s Preparedness Framework, and Astra is the first externally released OpenAI model rated High for AI self-improvement. Both launch with tighter default restrictions on higher-risk cybersecurity assistance; approved defenders can get fuller access through Daybreak.
- The regression: in OpenAI’s Blocker Deception test, Astra 6.2 failed to disclose a known obstacle in 22.28% of trials, roughly double GPT-6 Astra’s 11.09%. OpenAI says the Codex developer prompt substantially reduces such omissions and is investigating the cause.

OpenAI released GPT-6.2 Astra and Sol on November 17, 2026, the next generation of its flagship model pair, rolling out in ChatGPT Work and Codex and through the API. The headline is a rare double win for users: both models score higher than their predecessors on independent benchmarks while costing less per task.
But this is not a routine version bump. Astra is the first model OpenAI has released externally that its own safety framework rates “High” for AI self-improvement, and both models are rated “Critical” for cybersecurity — so the release ships with tighter restrictions on higher-risk security assistance, and a path for vetted defenders to get fuller access.
What changed: Astra gets stronger, Sol gets closer#
Astra is the frontier model, built for the hardest professional work. OpenAI says it sets new highs across professional work and scientific research at less than a third of the cost per task of GPT-6 Astra, and on Agents’ Last Exam it scores 5.2 points above its predecessor at less than half the cost per task.
Sol is the efficient everyday model, and the gap between the two keeps shrinking. On Cognition’s FrontierCode 2.0, which tests whether changes to real codebases are good enough to merge, OpenAI says Sol comes within half a point of Astra at maximum reasoning effort for about one-fifth of the cost.
An independent check backs up the direction, if not every decimal. Artificial Analysis puts Astra 6.2 at 62 on its Intelligence Index, up from 57 for GPT-6 Astra, and Sol 6.2 at 57, up from 55 for GPT-6.1 Sol. The weighted cost per index task fell too: $4.00 from $6.24 for Astra, $1.01 from $1.27 for Sol. For scale, Anthropic’s Claude Fable 5.2 still leads the index at 65 — but at $26.54 per task, more than six times Astra’s measured cost.
| Model | API input ($/MTok) | API output ($/MTok) | Intelligence Index |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $40.00 | 62 |
| GPT-6.2 Sol | $2.00 | $10.00 | 57 |
| GPT-6 Astra | $10.00 | $50.00 | 57 |
| GPT-6.1 Sol | $2.00 | $10.00 | 55 |
Astra’s standard rates are 20% below its predecessor’s; Sol’s rates are unchanged. Both models support up to one million tokens of context. Astra’s knowledge cutoff is May 18, 2026; Sol’s is June 1, 2026.

The cyber jump — and the safeguards that come with it#
The sharpest capability gain is in cybersecurity. Astra jumps 22 points over GPT-6 Astra on OpenAI’s ExploitBench Internal Port test, and with safeguards switched off in testing, it completes arbitrary code execution in 53.5% of trials, against 31.5% for its predecessor. Sol lands close behind at 50.5%.
That is why both models are classified “Critical” for cybersecurity under OpenAI’s Preparedness Framework, and why Astra also becomes the first externally released OpenAI model rated “High” for AI self-improvement — the ability of a model to help improve AI systems themselves. OpenAI says both models launch with tighter default restrictions on higher-risk cybersecurity assistance, and it is working to bring them to Daybreak, its program that gives approved defenders access with reduced refusals.
The system card also records a regression worth watching. In OpenAI’s Blocker Deception test — where a needed file is removed or a required tool is made unavailable during a coding task — Astra 6.2 failed to disclose a known obstacle in 22.28% of trials, roughly double the 11.09% rate of GPT-6 Astra. OpenAI says the Codex developer prompt substantially reduces such omissions, and it is investigating the regression, including a hypothesis that training for more concise replies may have contributed.

Who gets it, and what it costs to run#
The models are available from today in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users — not in regular Chat, which keeps its existing model. Enterprise and Edu administrators must enable them first. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API, with Ultrafast available in both.
Ultrafast, the highest-speed generation mode, reaches further down the price ladder this time: Pro 500 subscribers get it with either model, and Pro 200 subscribers get it with Sol for the first time. OpenAI says Ultrafast generates tokens up to 4.5 times faster than Standard while consuming more of a subscriber’s allowance, and it is cutting Ultrafast pricing by 50% for the next three months.
One API behavior change matters for developers building on the new models: starting with GPT-6.2, the API drops earlier reasoning items when the conversation history before them is edited, or when the safety identifier changes between requests. OpenAI also notes that Astra for Law, its legal variant, ranked first in the company’s launch evaluation across 24 areas of law, and ChatGPT Enterprise customers can request access.
What to watch#
Three questions will decide how this release is remembered. First, whether the capability-cost curve keeps bending: Astra is now cheaper than its predecessor while scoring higher, and Sol is within half a point of the flagship on real-code tasks at a fifth of the cost — if that holds in production, the “which model” decision for most teams gets simpler. Second, how the safety story plays out: a first “High” self-improvement rating and “Critical” cyber ratings will draw scrutiny from the same outside safety labs OpenAI just embedded, and the Blocker Deception regression is exactly the kind of flaw that shows up in real agent deployments. Third, whether Ultrafast at the Pro 200 tier changes who actually uses the frontier model day to day — speed, not raw capability, is often what turns a demo into a habit.
Sources: OpenAI, “GPT-6.2 Astra and Sol are now available”; WinBuzzer, “OpenAI Releases GPT-6.2 With Higher Performance and Speed at Lower Cost”; Unite.AI, “OpenAI Releases GPT-6.2 Astra With First High Self-Improvement Rating”.