OpenAI released GPT‑6.2 Astra and GPT‑6.2 Sol, available via API as gpt-6.2-astra and gpt-6.2-sol. Astra pricing dropped 20% versus GPT‑6 Astra to $8 per million input tokens and $40 per million output tokens, matching Claude Fable 5.2’s rate; cached input pricing fell sharply to $0.40 per million from $1 per million. Sol pricing remains $2/$10 per million, half of GPT‑5.6 Sol’s $4/$20. Simon Willison added support in LLM 0.38 and benchmarked both models at “Ultra” effort in the Codex CLI using a Three.js prompt requiring Yonkers to morph into a mech stomping across Manhattan, Queens, and Brooklyn with OpenStreetMap data. Neither model matched the morphing quality of Claude Fable 5.2 or Gemini 4 Carbon. Willison generated side‑by‑side comparison videos (2000×2000, 20 s) from each model’s output. Early feedback is mixed: Astra 6.2 trails Fable 5.2 on benchmarks and exhibits a confusing communication style. Theo Browne’s pre‑release review noted that during a two‑hour “Outer Wilds level” task, Astra spawned subagents effectively but then blocked on an opaque internal detail — asking whether “r24’s reviewer writes its clips to 24_g521.mp4 or 24_p11.mp4” — illustrating the model’s tendency to surface inscrutable implementation questions.
Read original
simonwillison