AI NewsModels & agentsAnnouncement

OpenAI released GPT-6.2 Astra and Sol with cheaper tokens and new scores

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol. Both have cheaper tokens and new top scores on Agents' Last Exam.

AI News

Editorial3 min read

LinkedInX
Product card for GPT-6.2 Astra and Sol on openai.com

Image: OpenAI

Why it mattersA team paying for GPT-6 Astra can keep its prompts and pay 20 percent less per task. Cached input is also 60 percent cheaper.

A team paying for GPT-6 Astra can move to a newer model and pay less for the same work. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Monday. The company puts both models into ChatGPT Work, Codex and the API. OpenAI says both models return answers in fewer output tokens. One paid plan gets more work done on the same quota.

GPT-6.2 Astra is the harder-task model. GPT-6.2 Sol is the everyday one. OpenAI says GPT-6.2 Astra sets a new high on Agents' Last Exam. That test measures long multistep work. OpenAI says it scored 5.2 points above GPT-6 Astra at less than a third of the cost. It also scored 5.5 points above Anthropic's Claude Fable 5.2, at 5 percent of the cost. These figures are OpenAI's own. The company notes that Fable 5.2's leaderboard number leaves out nine tasks where a fallback model ran.

What changed for buyers of API tokens

API prices for GPT-6.2 Astra are $8 per million input tokens and $40 per million output tokens. Cached input is $0.40 per million. OpenAI says that is 20 percent below GPT-6 Astra on input and output. Cached input is 60 percent cheaper. GPT-6.2 Sol keeps the GPT-6.1 Sol rate. That is $2 for input, $0.10 for cached input and $10 for output, per million tokens. OpenAI says teams should keep prompt prefixes the same and add new instructions at the end. This way the cache is reused and the saving is bigger.

Both models are available today to Plus, Pro, Enterprise and Edu users of ChatGPT Work and Codex. Pro 500 subscribers get Ultrafast on both. Pro 200 subscribers get Ultrafast on GPT-6.2 Sol for the first time. Enterprise and Edu users wait for a workspace admin to turn it on. Developers can call gpt-6.2-astra and gpt-6.2-sol through the API. Ultrafast is on both at launch.

Named outside tests

OpenAI quotes four customer figures as the customers' own. Legora's CTO Jacob Lauritzen says GPT-6.2 Astra used 62 percent fewer tokens than GPT-6.1 Sol on complex legal-agent tasks. Niko Grupen is Harvey's Head of Applied Research. He says GPT-6.2 Astra completed three times as many tasks on Harvey's Legal Agent Benchmark as GPT-6 Astra did. He also says it used 34.5 percent fewer output tokens. ClickUp's Head of AI Jay Hack says GPT-6.2 Astra beat ClickUp Brain's production model in blind testing. He says it used about 40 percent fewer tokens and had 30 percent fewer review rounds. IMC's Marquis Wong says GPT-6.2 Astra's writing and diagrams were easier to read than GPT-6 Astra's.

OpenAI also quotes its own benchmark numbers. It says GPT-6.2 Astra scores 9 points higher than GPT-6 Astra on Terminal-Bench-Science 0.1. It also says the cost is under a third as much. The company says it leads on GDPval v2 by 6.2 points over GPT-6 Astra, with 22 percent fewer output tokens. It says GPT-6.2 Astra more than doubles Claude Fable 5.2's pass rate on Harvey LAB-AA v1.1. It also says the cost is about one sixth. Treat every one of these numbers as OpenAI's own until a third party reports the same test.

A team's real cost question is per-task cost on its own work. A vendor number on a vendor's own prompt does not answer it. Replay a week of one real agent task through GPT-6.2 Sol and your current model. Score the outputs with your own eval. Read the total token bill. Fewer tokens per task is the one gain OpenAI keeps naming. It is also the one that is easy to check on real traffic.

Source

OpenAI: Introducing GPT-6.2 Astra and Sol.

SourceOpenAI

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project