OpenAI launches GPT-6.2 Astra and Sol

OpenAI has released GPT-6.2 Astra and Sol, and the most interesting claim in the announcement is about the people who will use Astra inside OpenAI itself. The company estimates that, compared with what its researchers had in 2024, giving them Astra is roughly equivalent to giving each of them a highly capable research engineer as an assistant.

Astra is the first model OpenAI has released publicly to earn a “High” rating for AI self-improvement under its own safety framework. Sol falls below that threshold. Now, “self-improvement” might have you picturing ChatGPT disappearing into a server room on Friday and emerging as GPT-7 on Monday. The tests behind that rating tell a more interesting story about where researchers actually spend their time.

Take a failed experiment. Somebody has to figure out what went wrong before the lab burns through another day of computing time. OpenAI tested the models on 41 bugs that had taken experienced researchers anywhere from hours to days to diagnose, along with six related auditing tasks. Astra earned about 82 percent of the available points. The graders gave partial credit for an incomplete diagnosis, so it would be a stretch to say Astra solved 82 percent of the bugs.

Another test gave Astra two hours and a single NVIDIA H100 GPU to improve the code used to train a small language model. Its changes cut the time needed to reach the target by nearly 43 percent on average. The best human solution cut it by more than 72 percent. I’d happily take the savings Astra found, and I’d also want to know what the human spotted that it missed.

Choosing what to investigate in the first place proved harder. Research nonprofit Epoch AI gave Astra open-ended assignments drawn from its own work and saw little improvement over the previous Astra on the tasks it tested. The model still struggled to design good experiments. Its writeups were somewhat better, although Epoch said they fell short of its standards.

There’s a more immediate concern for people using these models in Codex. According to OpenAI’s system card, both were more likely than GPT-6 Astra to leave out an obstacle they had encountered, such as a missing file or a broken search tool. OpenAI says changing the instructions in Codex reduced those omissions. If I ask a coding agent to fix a bug, I’d like to hear that it couldn’t open my file before I spend an afternoon testing the supposed fix.

For developers choosing between the two models, Sol makes a pretty good case for itself. On Cognition’s FrontierCode 2.0, which tests whether changes to a real codebase meet a maintainer’s standards for merging, Sol finished within half a point of Astra. Both ran at maximum reasoning effort, and Sol did the work for about a fifth of the cost.

Both models are available through the API and in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. At standard API rates, Sol costs $2 per million input tokens and $10 per million output tokens. Astra costs four times as much at $8 and $40, respectively. If I were paying by the token to work on an app, I’d start with Sol.

What I’d like to see from OpenAI is a number it can get from its own lab. How much time passes between a researcher proposing an experiment and getting a result the team trusts? Count the hours people spend checking Astra’s work and chasing down its mistakes, too. That would tell us how much sooner OpenAI can actually build its next model.

☕

Support independent tech journalism

NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.

Support NERDS.xyz
Written by

Brian Fagioli ✔

Technology journalist and founder of NERDS.xyz

Brian Fagioli is a technology journalist and founder of NERDS.xyz. A former BetaNews writer, he has spent over a decade covering Linux, hardware, software, cybersecurity, and AI with a no nonsense approach for real nerds.

Leave a Comment