Skip to content
Sir CompareComparisons with evidenceYouTubeTelegram
United States · USD

News

GPT-6.2 Astra and Sol: prices, API names, and what Simon Willison saw

Updated

Illustrative photo: AI Tools & SaaS: Code on computer monitor (Unsplash)
Illustrative photo

Two models, two API names

OpenAI has introduced GPT-6.2 Astra, described as its new frontier model for the hardest projects, and GPT-6.2 Sol, an efficient, versatile model for everyday work. Both are available in the API under the names gpt-6.2-astra and gpt-6.2-sol.

Simon Willison writes that he shipped LLM 0.38 with support for the new models.

The products in this article

Astra's price drops to $8 and $40 per million tokens

Astra 6.2 is priced at 80% of what 6 Astra cost: $8 per million input tokens and $40 per million output tokens. Willison notes that this matches the price of Claude Fable 5.2, which was also 80% of its predecessor. 6 Astra and Fable 5 were both $10/$50.

The cached price is a steeper cut. It sits at $0.40 per million against $1 per million for the older model. Willison calls this good news for longer running agentic sessions.

Sol holds its price at $2 and $10

GPT-6.2 Sol keeps the same $2/$10 price as 6.1 Sol and 6 Sol. Those two were half the price of GPT-5.6 Sol, which sat at $4/$20.

So the cheaper everyday model doesn't move on cost with this release. If you were already paying $2 input and $10 output per million tokens for Sol, that number stays where it was.

The pelican test and a three.js prompt

Willison ran his usual pelican-riding-a-bicycle prompt on both models and published a full comparison grid, including how the high, xhigh and max versions stack up against earlier Sol and Astra releases.

He also ran a recent benchmark prompt at "Ultra" effort levels in the Codex CLI: a three.js animation of Yonkers standing up as a big mech and stomping around Manhattan, Queens and Brooklyn using OpenStreetMap data. In his words, neither model morphed Yonkers into a big mech well enough, at least compared with Claude Fable 5.2 and Gemini 4 Carbon.

Mixed early reaction and a subagent complaint

Willison describes the buzz around the new models as mixed. Astra 6.2 isn't scoring quite as high as Fable 5.2 on various benchmarks, and he has seen complaints that its communication style can be confusing.

He points to Theo Browne's review, written as a feedback letter to OpenAI during pre-release. Browne describes a model that pulls in many subagents on long tasks and manages them well, then stops to ask about an internal implementation detail the user couldn't possibly understand. In his example, a request to make a new level for Outer Wilds was blocked two hours in by a question about which file a reviewer's clips get written to.

Related coverage

Source: Simon Willison · LLM notes

Hardware to try AI at home

Affiliate links: we may earn a commission from qualifying purchases, and it never changes our data or rankings. They open Amazon's search results: prices and availability are shown there.

Video

What changed

  • Page updated

How I put products in order

Sir Compare’s letter

Updates on AI tools & SaaS? I’ll send them your way.

Pick what you care about: I only write when something new comes out — a guide, a news story, a verified price that drops.

Step 1 of 2

What are you into?

Pick as many as you like. Pick none and you get what’s new in every section.

What you want to receive
  • Only what you pick
  • Never an empty email
  • One click to leave

Your assistant

Open full page →