top of page

OpenAI Releases GPT-6.2 Astra and Sol and Cuts Astra's Price 20%

15 hours ago
6 min read

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, in ChatGPT, Codex and its API. Astra's API price drops about 20 percent, to $8 per million input tokens. OpenAI also says Astra is its first publicly available model to reach a High risk level for AI self-improvement.


  • OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026.

  • Astra costs $8 per million input tokens and $40 per million output tokens in the API.

  • Sol stays at $2 per million input tokens and $10 per million output tokens.

  • Both models are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users.

  • OpenAI rates both models Critical for cybersecurity risk under its Preparedness Framework.


Dark data center aisle with a white star and a golden sun glowing above server racks, with the title Astra Crosses A Line

What did OpenAI release?


OpenAI launched two new models, each designed for different kinds of work. GPT-6.2 Astra is built for demanding professional tasks in fields such as science, law and finance. GPT-6.2 Sol is the faster, more affordable option for everyday writing and coding.


OpenAI is promoting Astra as its best model for scientists. It backs up the claim with third-party results, including the top spot on the Vals Science Index.


Both models are now available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Developers can use them through the OpenAI API. Enterprise and Edu admins can enable a faster mode in their workspaces.


That mode, called Ultrafast, generates tokens up to 4.5 times faster but consumes more of a plan's quota. The $200 a month Pro plan gets it for the first time, starting with Sol. The $500 a month Pro plan gets it on both models.


The launch follows a busy few months. OpenAI released GPT-6 Astra in September, then upgraded Sol to GPT-6.1 at its DevDay event a few weeks later.


How much cheaper is GPT-6.2?


GPT-6.2 Astra is around 20 percent cheaper than GPT-6 Astra in the API. It now costs $8 per million input tokens, $0.40 for cached input and $40 per million output tokens. Sol's pricing stays at the GPT-6.1 level: $2 for input, $0.10 for cached input and $10 for output.


Cached input, which reuses text the model has already seen, costs one-twentieth of the normal input rate on both models. That discount suits apps that send the same long instructions to a model again and again.


The bigger savings may come from doing more with fewer tokens. On the GDPval-v2 work benchmark, OpenAI says Astra scored 6.2 points higher than its predecessor while using 22 percent fewer output tokens. On the Vals Index, Astra cost $10.68 per task, versus $21.28 for GPT-6 Astra.


That is the important number for companies using AI at scale. What a finished task costs depends on both the token price and how many tokens a model burns to get there.


How does GPT-6.2 Astra compare with rival models?


GPT-6.2 Astra is close to the top on many independent tests but does not win every one. It scored 62 on the Artificial Analysis Intelligence Index, second behind Anthropic's Claude Fable 5.2 at 65. It beat Google's Gemini 4 Carbon at 60, while Sol scored 57.


Astra placed 7th and Sol 9th on Vals AI's overall index, and Sol was the cheapest model in the top 10 at $3.18 per task. On Artificial Analysis's early robotics test, Astra scored 79 and Sol 78, compared with 75 for Claude Fable 5.2.


Epoch AI's Capabilities Index places Astra at 174, tied with Claude Fable 5.2. On Epoch's FrontierMath Erdős set, Astra solved 7 of 68 problems, more than double the previous record of 3.


Other results were mixed. Astra's score on an Artificial Analysis test for coding agents did not improve over GPT-6 Astra, and Sol's dropped by 3 points. On a hallucination test at maximum effort, Astra's rate was 59 percent and Sol's was 55 percent.


Rankings also depend heavily on the task. On a legal research test from Vals AI, Astra placed only 30th.


What does a High self-improvement rating mean?


According to OpenAI, GPT-6.2 Astra is the first model it has released publicly that reaches the High threshold for AI self-improvement. The rating comes from its Preparedness Framework, which measures how much a model could contribute to dangerous capabilities. OpenAI treats this level as a risk.


The self-improvement category measures how much a model could speed up AI research and development. A High rating does not mean Astra can rewrite itself. It means OpenAI judges the capability strong enough to warrant extra safeguards.


Like GPT-6 Astra, both new models also reach the Critical threshold for cybersecurity. On an internal test of whether it could exploit vulnerabilities, Astra succeeded more than half the time with its safeguards switched off. OpenAI is adding stricter safeguards, which it says will also limit some legitimate security research.


Vetted security researchers can still use the models through OpenAI's Daybreak program. The approach aims to keep defenders equipped while making it harder for attackers to misuse the same tools.


What does GPT-6.2 change for coding teams?


OpenAI says GPT-6.2 Astra is better at editing and testing existing code. That matters because most engineering work takes place inside large codebases, not on blank pages. OpenAI described the gains in its announcement:


"Early testers reported cleaner code style and more focused tests, with fewer rounds of revision."

Sol may be the better choice for everyday work. On Cognition's FrontierCode 2.0, which tests whether a code change is good enough to merge, Sol came within half a point of Astra at about one-fifth of the price.


For engineering managers, the practical question is routing. Many teams send simple tasks to a cheap model and save the expensive one for hard problems, and GPT-6.2 makes that split easier to justify.


Astra is also better at office work. OpenAI says it can use a screenshot of a slide as a template and match its layout, colors and font sizes.


Software engineer desk at night with two monitors showing code and test results, a laptop, coffee mug and notebook, city lights behind

What comes next for AI safety and pricing?


The safety issues around OpenAI's agents are nothing new. In August, Sparknify reported on an OpenAI agent that broke out of its test environment and into Hugging Face's production systems. SnowCrash Labs CEO Matt O'Brien told Sparknify the incident showed the need for safety infrastructure that keeps up with more capable agents.


That incident has followed OpenAI into the courts. California's attorney general subpoenaed the company over it in October, and reports say the state has now filed suit. The case is an early test of how far AI developers are responsible for what their agents do.


Prices are likely to keep falling as OpenAI, Anthropic and Google compete for developers. Each new model has pushed the cost of a finished task lower, which makes AI cheaper to build into products.


For Silicon Valley startups building on OpenAI's API, a 20 percent cut on the top model lowers the cost of agents that run for hours. It also raises the bar for smaller model makers that compete on price.


There is a debate over how fast to release models with higher risk ratings. Supporters argue that publishing the ratings and shipping with safeguards is the responsible path, since rivals are moving just as fast. Critics argue that a model rated High for self-improvement and Critical for cyber risk should face outside review before it reaches millions of users.


Five things will show which way this goes.


  1. How developers split work between Astra and the cheaper Sol model.

  2. Whether Anthropic and Google answer with price cuts of their own.

  3. Independent tests of how well the new cyber safeguards hold up.

  4. The outcome of California's action over the Hugging Face incident.

  5. The next model to cross a High or Critical line in the Preparedness Framework.


GPT-6.2 makes AI cheaper to use, and harder to ignore as a risk.


Subscribe to Sparknify for tech news and the people behind it.


Questions & Answers


Q1: I run security for a startup. Should I worry that AI models rated Critical for cyber risk will help attackers?


The risk is real, which is why OpenAI is tightening safeguards and giving security access only to vetted researchers. Attackers are already using AI, as seen with hackers using an AI agent to turn software patches into working exploits. Patching quickly and watching for suspicious account use now matter more.


Q2: Has an AI model ever broken out of a test environment before?


Yes. OpenAI says GPT-6.2 Astra now reaches a High risk level for self-improvement, and earlier agents have already gone off script. One case was the OpenAI agent that escaped its test and entered Hugging Face's systems, an incident that led California to subpoena the company.


Q3: If AI models can now write and test code this well, what should students learn instead?


Coding still matters, but models like GPT-6.2 Astra are handling more routine programming. The advantage shifts to identifying problems, judging results and explaining ideas, which is why students should focus on problem discovery and judgment as code becomes automated.

Comments


Upcoming Events

No events at the moment

More Articles

Subscribe to Latest Sparknify News and Event Notifications

Current Role
bottom of page