OpenAI's GPT-6.2 Astra and Sol target long agent workflows

OpenAI says the new models finish multistep professional work with fewer tokens at lower cost. Independent testers found no gain on a coding-agent index, and some early reviewers were harsh.

Illustration of a blue four-pointed star and a golden sun on a dark blue background, linked by a glowing chain of task icons that runs from play, grid and summit icons to search, checkmark and bar-chart icons. Image: AgenticFeed illustration

OpenAI on Tuesday released GPT-6.2 Astra and GPT-6.2 Sol, two models it says complete long, multistep professional work with fewer output tokens and at a lower cost per task than their predecessors. The company announced the launch on X and in a blog post.

GPT-6.2 Astra is OpenAI's new frontier model "for your hardest projects," while GPT-6.2 Sol is pitched as an efficient model for everyday work, according to the post. Both are rolling out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.2-astra and gpt-6.2-sol.

Built for long agent runs

OpenAI framed much of the launch around agentic tasks. On Agents' Last Exam, a benchmark of long-horizon professional workflows, OpenAI said GPT-6.2 Astra scored 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Anthropic's Claude Fable 5.2 at 5% of the cost.

The company said Astra has "better judgment about how to execute a task," including knowing when to use plugins versus computer use. It also said Astra's computer use is precise enough to work with individual parts in CAD and simulation tools.

Customer testimonials in the post focused on efficiency. Basis, one of the testers quoted by OpenAI, said its average time for long-horizon accounting work fell by 50%. Harvey said Astra completed more than three times as many Legal Agent Benchmark tasks as GPT-6 Astra with 34.5% fewer output tokens.

Pricing

In the API, GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens, according to OpenAI. The company said that is 20% below GPT-6 Astra on input and output, with a 60% cut on cached input. GPT-6.2 Sol keeps GPT-6.1 Sol's rates of $2 for input and $10 for output.

OpenAI also said it is bringing its Ultrafast option to Pro 200 subscribers for the first time, starting with GPT-6.2 Sol. Pro 500 subscribers can use Ultrafast with either model, according to the company.

Independent numbers are mixed

Artificial Analysis, which runs independent benchmarks, said GPT-6.2 Astra now ranks second on its Intelligence Index, behind only Claude Fable 5.2, after a 5-point gain over GPT-6 Astra. It said OpenAI models now occupy most of its cost-per-task frontier.

The firm also reported no improvement on its Coding Agent Index, run in the Codex harness at maximum effort. Astra was flat against GPT-6 Astra, and Sol lost 3 points against GPT-6.1 Sol, driven by a 14-point drop on one question-answering test where it was "more likely than its predecessor to omit key facts," Artificial Analysis said.

Some early testers were blunt. Developer Theo wrote on X that Astra was "the worst model OpenAI has ever released," and investor Matt Shumer said the models were "essentially unusable" in his testing. Tibo Sottiaux, whose X profile lists him as working on Codex and ChatGPT at OpenAI, replied that some reviewers had access weeks earlier and that the company made "a lot of changes in the last 3-4 days" before launch.

Safety flags

OpenAI said both models meet the Critical threshold for cybersecurity under its Preparedness Framework, and that it launched them with tighter safeguards for higher-risk security requests. The company said GPT-6.2 Astra is its first externally released model to reach the High threshold for AI self-improvement.

The company also disclosed an agent-specific weakness. In two alignment tests, the models were more likely to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool, OpenAI said. It added instructions in Codex that it said "substantially reduce these omissions" and said it is still investigating the behavior.

OpenAI said it will work to bring the GPT-6.2 models to Daybreak, its program that gives vetted cyber defenders access to models with fewer refusals. Defenders there currently use GPT-6 Sol, an alpha of GPT-6 Astra and GPT-6 Cyber, according to the post.

Source: OpenAI, “Introducing GPT-6.2 Astra and Sol”

Additional sources: OpenAI on X, “GPT-6.2 Astra and Sol are now available”; Artificial Analysis on X, “GPT-6.2 Astra and Sol in the Artificial Analysis Intelligence Index”; Tibo Sottiaux on X, “On early reviews of GPT-6.2”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 19 Nov, 2026

MCP Dev Summit Nairobi

The Linux Foundation’s Nov. 19 and 20 summit brings Model Context Protocol developers to Kenya, with keynotes from OpenAI, GitHub and the Gates Foundation.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.