OpenAI’s newest frontier model can take reference photos of a building and turn them into a paper lamp design you build by hand. In its launch post for GPT-6.2 Astra and Sol, OpenAI shows Astra turning the scalloped balconies of Chicago’s Marina City towers into a lamp, via CAD revisions. That is OpenAI’s own demo. One early outside index ranks Astra second, behind Anthropic’s Claude Fable 5.2.

Both models launched on 17 November, per OpenAI’s ChatGPT help page. Astra is for the hardest projects; Sol is the cheaper everyday model.

What the two models are for

OpenAI says Astra’s slides, spreadsheets and documents “read less like AI” and follow your templates more closely. Asked to pitch animal-themed merch in OpenAI’s public slide template, it returned a deck styled as an illustrated field guide, with vector artwork and proposed sizes for each design.

Other demos include a playable racing game and an animated assembly guide for a radio-controlled buggy. Sol is aimed at code reviews, writing and multistep tasks, and OpenAI says many of its employees use it as their default.

Where you can try it, and what it costs

The models live in ChatGPT Work and Codex, OpenAI’s coding agent. An OpenAI staff post in its developer forum says the rollout does not add them to regular Chat. The help page lists Plus, Pro, Business, Enterprise and Edu plans, with new Work chats on consumer plans starting on Sol at Light effort.

In the API, Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra. Cached input drops to $0.40, from $1.00 according to Artificial Analysis. Sol stays at $2 and $10.

That per-token price equals Claude Fable 5.2’s, which Anthropic lists at $8 and $40, with cheaper cached reads at $0.20. Much cheaper models exist too, such as Alibaba’s Qwen4-Max, which opened to API customers this week.

OpenAI’s claims versus outside results

OpenAI says Astra beats Fable 5.2 by 5.5 points on Agents’ Last Exam, a test of long professional workflows, at 5% of the cost. A footnote adds that Fable’s result leaves out nine tasks where it handed off to a fallback model. OpenAI also says Astra more than doubles Fable’s pass rate on Harvey’s legal test.

Outside testers are more mixed. Artificial Analysis gives Astra 62 on its Intelligence Index, second behind Fable 5.2, and Sol 57. On its coding agent index, Astra scored no better than its predecessor and Sol slipped.

It also measured hallucination rates of 59% for Astra and 55% for Sol at max effort. That counts how often a model guesses on hard factual questions instead of admitting it does not know.

Epoch AI scored Astra level with Fable 5.2 on its overall capabilities index, The Decoder reports, but behind Fable on Epoch’s own research tasks. Vals AI ranked Astra 30th on its legal research benchmark, a different test from the one OpenAI cited. Our guide to how the big AI leaderboards differ explains why rankings disagree.

The safety notes

OpenAI calls Astra its first released model to reach its “High” threshold for AI self-improvement. It likens that to giving each of its researchers an extra, highly capable research engineer. In two alignment tests, both models were more likely to skip mentioning an obstacle, such as a missing file. OpenAI says new Codex instructions substantially reduce this.

What to do this week

On a paid ChatGPT plan, open Work, pick Astra and give it a real template and report from your job, then check every figure it writes. On the API, run your own prompts on both models and compare cost per finished task, not the per-token price.

Sources

Lars Strother

Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.