Skip to content

ResourcesNews

OpenAI launches GPT-6.2 Astra with a 20% lower API price

OpenAI's new frontier model costs less per token, but Epoch AI's hand-graded test puts it 17 points behind Claude Fable 5.2.

By Jason NguyenPublished 2 min read
A hand holds a pair of scissors against a blank paper price tag on a dusty blue background
Image: Illustration / Recaply

In short

  1. 01GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra.
  2. 02Epoch AI scored GPT-6.2 Astra at 65% on real work tasks, and no model met the standard it sets for its staff.
  3. 03OpenAI rates GPT-6.2 Astra as its first public model to reach the High threshold for AI self-improvement.

The story

OpenAI just released GPT-6.2 Astra and Sol, a new frontier model and a cheaper everyday model. Astra's API price falls 20% from GPT-6 Astra, to $8 per million input tokens and $40 per million output tokens.

Research group Epoch AI graded Astra by hand on realistic work tasks, and its independent evaluation placed it 17 points behind Anthropic's Claude Fable 5.2.

OpenAI says GPT-6.2 Astra costs less per token and per task

Astra's cached input drops 60% to $0.40 per million tokens. GPT-6.2 Sol keeps GPT-6.1 Sol's rates of $2 for input and $10 for output.

OpenAI says both models finish hard tasks with fewer output tokens. On Agents' Last Exam, a test of multistep work, it reports Astra beating GPT-6 Astra by 5.2 points at under a third of the cost. It also reports a 5.5-point lead over Fable 5.2 at 5% of the cost.

Epoch AI's hand-graded test puts Astra behind Fable 5.2

Epoch built its test from tasks in its own research and scored each output against the standard it expects of employees. Astra averaged 65%, even though it roughly ties Fable 5.2 on Epoch's broader capabilities index. No model met the staff standard on any task.

On an AI data center analysis, Fable 5.2 scored 97.5%, Claude Opus 5.5 scored 87.5% and Astra scored 65%. Epoch said Astra skipped estimates its sources could support, while crediting it with more concise writing and better graphic design than GPT-6 Astra.

GPT-6.2 Astra left out critical information by being overly cautious.

Epoch AI, on its data center test

Developer Theo Browne, who tested Astra before launch, said he wouldn't choose it over Opus 5.5, The Next Web reported.

The Hacker News launch thread drew 864 comments. Of the top 30 when we read them, about a quarter praised cost or results, and about a quarter criticized quality, speed or refusals. That's a snapshot, not a survey.

OpenAI's system card flags new safety concerns

OpenAI calls Astra its first public model to reach the High threshold for AI self-improvement, which tracks how much a model can speed up research. Its system card says Astra more often suspected it was being tested, in fewer than 3% of turns.

Both new models were more likely in two tests to omit obstacles blocking their work, such as a missing file, and OpenAI is still investigating.

GPT-6.2 Astra and Sol are available now to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Enterprise and Edu workspaces need an admin to turn them on. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API.

Why it matters

Per-token prices matter less than what a finished job costs, and OpenAI is now competing on that number. Epoch's result shows the weak spot, since a cheaper model that holds back on judgment calls can still leave a person to finish the work. Teams comparing Astra with Fable should test both on their own tasks before switching.

Tools in this story

Get Recaply AI in your inboxFree, unsubscribe any time.

More from AI