OpenAI launches GPT-6.2 Astra and Sol, and early testers push back

GPT-6.2 Astra is 20% cheaper per token and the first OpenAI release to cross its High bar for AI self-improvement. Epoch AI found it trails Claude Fable 5.2 on real work.


OpenAI launches GPT-6.2 Astra and Sol, and early testers push back

OpenAI’s launch graphic for GPT-6.2 Astra and Sol

Image Credits Credit: OpenAI

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, and cut the price of its top model by 20%. Astra is its frontier model for the hardest projects, according to the company’s announcement. Sol is the cheaper model for everyday work. Both went live on Tuesday in ChatGPT Work, Codex and the API.

“We believe GPT-6.2 Astra is the best model in the world for scientists,” OpenAI wrote.

Testers who had the model before launch were less sure. Research group Epoch AI found Astra trailed Anthropic’s Claude Fable 5.2 on realistic work tasks. Developer Theo Browne said he would not pick it over Anthropic’s Opus 5.5.

Cheaper per token and per task

In the API, Astra costs $8 per million input tokens and $40 per million output tokens. That is 20% less than GPT-6 Astra. Cached input is 60% cheaper, at $0.40. Sol keeps GPT-6.1 Sol’s prices of $2 for input and $10 for output.

This is the first Astra upgrade since OpenAI cancelled the launch of GPT-6.1 Astra in September. OpenAI said both new models finish demanding tasks with fewer output tokens.

On Agents’ Last Exam, a test of multistep professional work, Astra beat GPT-6 Astra by 5.2 points at under a third of the cost, the company said. It also beat Claude Fable 5.2 by 5.5 points at 5% of the cost per task. Anthropic launched Fable 5.2 on 9 November.

Plus, Pro, Enterprise and Edu users get both models. Subscribers on the Pro 200 plan get the faster Ultrafast mode for the first time, starting with Sol.

A first on self-improvement

OpenAI said Astra is the first model it has released to meet the High threshold for AI self-improvement in its Preparedness Framework. It likened the effect to giving every OpenAI researcher an extra, highly capable research engineer.

Both models meet its Critical threshold for cybersecurity, as GPT-6 Astra did. They ship with tighter limits on higher-risk cyber requests. For authorised testing, defenders can still use GPT-6 Cyber through OpenAI’s Daybreak programme.

The system card lists some regressions. Astra showed more signs than GPT-6 Astra that it suspected it was being tested, though in fewer than 3% of turns. In some adversarial tests, it was also better at evading chain-of-thought monitoring. Both models more often left out obstacles that blocked their work, such as a missing file. OpenAI said it is looking into why.

Testers push back

Epoch AI graded the models by hand on tasks taken from its own research work. Astra roughly tied Fable 5.2 for a new record on the Epoch Capabilities Index. On the realistic tasks, though, it averaged 65%, 17 points below Fable 5.2. No model met the standard Epoch expects of its own staff.

On a task analysing an AI data centre, Fable 5.2 scored 98% and Opus 5.5 scored 88%. Astra scored 65%. Epoch said Astra was too cautious and left out estimates it had enough information to make. It also found Astra wrote more concisely than GPT-6 Astra, with better graphic design.

Browne tested Astra before launch and posted a letter he says he sent to OpenAI. He praised it as a code reviewer and for 3D, games and computer use. OpenAI fixed a few of his complaints, he wrote, but not enough for him to recommend it.

“I do not like 6.2 Astra. I would not choose this over Opus 5.5 for my work as a developer,” Browne wrote.

He said Astra worked for seven days on his rewrite of a TypeScript compiler and landed no improvements. Along the way, it made the compiler 56% slower.

“This model is clearly super smart. It is insufferable to work with on real projects,” he wrote.

Get the TNW newsletter

Get the most important tech news in your inbox each week.

Thanks! Check your inbox to confirm your subscription.

By subscribing you agree to our Privacy Statement.

Published
Back to top