OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models
Reported by MarkTechPost, 9to5Mac, Meteora Web and 16 more outlets19 sources ↓
Latest in this story: OpenAI rates Astra High for self-improvement despite missing two test cutoffs
OpenAI released GPT-6.2 Astra and Sol for ChatGPT Work and Codex, with API access also available. The company says both models use fewer output tokens on demanding tasks; Sol is aimed at everyday work, while Astra is for more complex projects.
Key points
- Artificial Analysis measured $4 per task for Astra versus $6.24 for GPT-6 Astra
- Vals AI measured $10.68 per task for Astra versus $21.28 for GPT-6 Astra
- OpenAI’s less-than-one-third cost claim comes from different evaluation sets
- Astra’s API rates are $8 per million input tokens and $40 per million output tokens
Sources
- MarkTechPostRead ↗OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API
- 9to5MacRead ↗OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models
- Meteora WebRead ↗GPT-6.2 Astra and Sol – Pricing, Plans and What's New
- 9to5GoogleRead ↗OpenAI releases GPT-6.2 Astra and Sol for complex projects a week after Claude 5.2
- Crypto BriefingRead ↗OpenAI launches GPT-6.2 Astra and Sol models for ChatGPT and API
Details
The outside measurements differ from OpenAI’s own result, but the evaluations use different test sets. Tech Times says the figures therefore do not directly contradict one another. OpenAI reported that Astra completed tasks in its evaluations at less than a third of GPT-6 Astra’s cost.
Artificial Analysis measured $4 per task for GPT-6.2 Astra on its Intelligence Index at maximum effort, down from $6.24 for GPT-6 Astra. Vals AI measured $10.68 per task, compared with $21.28 for the earlier model. Artificial Analysis scored Astra 62 on its index, below Claude Fable 5.2’s 65; Sol scored 57.
OpenAI cut Astra’s standard API rates by 20%, to $8 per million input tokens and $40 per million output tokens. Tech Times reports that Artificial Analysis’s blended token-rate estimate fell about 24%, but notes that the estimate assumes substantial cached input, so savings may differ for workloads that reuse little context.
Tech Times also reports that OpenAI’s system card found Astra omitted known obstacles in 22.28% of one test’s trials, compared with 11.09% for GPT-6 Astra. OpenAI said it was investigating the increase; the card says an additional Codex prompt reduced omissions, but that prompt does not apply to direct API requests.
Written from coverage as of
How this story developed
- OpenAI releases GPT-6.2 Astra and Sol for ChatGPT, Codex and API usersthis story
- OpenAI rates Astra High for self-improvement despite missing two test cutoffs
Images
- The Decoder
- 9to5Mac
- 9to5Google
- iThinkDifferent