Enki

GPT-6.2 Astra launches at $8 in, $40 out, Sol at $2

By

Published

Reporting from OpenAI

More onOpenAI

GPT-6.2 Astra costs $8 in, $40 out per million tokens, 20% below GPT-6 Astra, while GPT-6.2 Sol holds at $2 and $10 in ChatGPT, Codex and the API.

What it means for founders

  • Rerun your cost model. On heavy Astra workloads, the 20 percent rate cut plus OpenAI's claim of fewer output tokens could shrink bills by more than the sticker price shows. Track cost per finished task.
  • Try Sol before paying for Astra. Sol nearly matches Astra on OpenAI's coding test, and Artificial Analysis saw no Astra coding gain over GPT-6 Astra.
  • Expect more cyber refusals. Security products built on exploit work may hit the new safeguards unless OpenAI follows through on plans to add these models to Daybreak, its access track for screened security teams.
  • Legal startups face a stronger default. Vals AI ranked the general models 30th and 52nd on legal research, while Astra for Law reached 6th. With any Enterprise customer able to request access through OpenAI's Trusted Access Program, you need an edge beyond research quality.

What we don't know yet

Daybreak access still has no date. It is not yet clear whether the looping early testers described persists in the models OpenAI shipped.

The story

OpenAI released two models on November 17, GPT-6.2 Astra for hard professional jobs and GPT-6.2 Sol as the cheaper daily model, and put both into ChatGPT Work, Codex and its API. OpenAI says Astra sets new highs on professional and scientific evaluations for under a third of GPT-6 Astra's cost per task.

GPT-6.2 API pricing per million tokens

The gpt-6.2-astra API model costs $8 for each million input tokens, $0.40 cached and $40 for each million output, 20 percent below GPT-6 Astra's $10 and $50. gpt-6.2-sol keeps the rates GPT-6.1 Sol launched with: $2 in, $10 out, $0.10 cached.

Plus, Pro, Enterprise and Edu accounts can pick both in ChatGPT Work and Codex, SiliconANGLE reported. Ultrafast mode, which OpenAI says generates tokens up to 4.5 times faster, now reaches the $200 Pro plan for Sol, according to 9to5Mac.

GPT-6.2 features: slides, computer use and coding

OpenAI says it trained Astra to follow the document conventions of law, finance, sales and accounting, and says it can copy the look of a slide from a screenshot. It also claims better computer use, demoing Astra animating a 3D car in Blender.

OpenAI says Sol trailed Astra by half a point on Cognition's FrontierCode 2.0 while costing roughly a fifth as much.

GPT-6.2 benchmarks: OpenAI's claims and outside tests

By OpenAI's count, Astra tops GPT-6 Astra by 5.2 points on Agents' Last Exam and leads Claude Fable 5.2 there by 5.5, at roughly 5 percent of Fable's cost. Outside results are mixed, The Decoder reported. Artificial Analysis ranks Astra second behind Fable 5.2 but found no coding gain over GPT-6 Astra, and Sol slipped 3 points below GPT-6.1 Sol on its coding agent index.

Some early testers saw the models loop or quit tasks early, Android Authority reported. OpenAI's Thibault Sottiaux said on X that late changes before launch may not have reached some reviewers, and that ease of use still needs work. OpenAI also says Astra is its first released model to hit its "High" level for AI self-improvement risk, and both models ship with tighter safeguards on risky cybersecurity requests.

Astra for Law 6.2 opens to ChatGPT Enterprise

In a separate post, OpenAI says any ChatGPT Enterprise customer can now request Astra for Law, its legal setup of GPT-6.2 Astra, including API access. On Harvey LAB-AA, 120 legal tasks spanning 24 legal fields, Astra for Law got full marks on 20.8 percent, ahead of base GPT-6.2 Astra at 15.0 percent and Muse Wonder Full at 13.1 percent, BigGo Finance reported.

Sources

Primary sources

Tools in this story

We may earn a commission if you sign up through our links. It never affects our ratings or which stories we cover.

ChatGPT

OpenAI's all-purpose AI assistant

9.2Editor’s scoreVisit ChatGPT
Codex

OpenAI's coding agent, included with ChatGPT

8.6Editor’s scoreVisit Codex

More in Models & Labs

How Enki covers newsCorrectionsReport an error