Wednesday, November 18, 2026

News

OpenAI Launches GPT-6.2 Astra and Sol, With Astra API Pricing 20% Below GPT-6 Astra

ModelsPatryk Raba
OpenAI Launches GPT-6.2 Astra and Sol, With Astra API Pricing 20% Below GPT-6 Astra
Fot. TechCrunch, Wikimedia Commons (CC BY 2.0)

OpenAI is releasing GPT-6.2 Astra for the hardest tasks and GPT-6.2 Sol for everyday work, in ChatGPT Work, Codex and the API. According to the company, Astra beats Claude Fable 5.2 on Agents' Last Exam by 5.5 points at about 5% of the cost, but that is a vendor claim, not an independent measurement.

Contents
  1. Two models, two roles
  2. What the benchmarks show
  3. Ultrafast and usage limits
  4. Safety caveats
  5. Context and what's next

On November 17, 2026, OpenAI introduced two new models: GPT-6.2 Astra, built for complex projects, and GPT-6.2 Sol, designed as a model for everyday tasks. Most of the performance figures come from the company itself.

Two models, two roles

According to OpenAI, Astra is the model for the hardest projects. It was trained to produce slides, spreadsheets and documents that follow professional conventions in law, sales, finance and accounting. It can also recreate a template from a file or a screenshot, including layout, colors and font size.

In Codex, Astra is meant to be better at making and testing changes in existing projects. According to OpenAI, early testers reported cleaner code style, more focused tests and fewer rounds of revision. Sol is the cheaper option for everyday coding and writing, and users are supposed to complete more tasks on the same plan than with GPT-6.1 Sol.

What the benchmarks show

OpenAI says Astra set a new record on Agents' Last Exam, which tests multistep professional work. It is said to beat GPT-6 Astra by 5.2 points at less than a third of the cost, and Claude Fable 5.2 by 5.5 points at about 5% of the cost. On Terminal-Bench-Science 0.1 it gained 9 percentage points over its predecessor, and on the Harvey LAB-AA legal-task test it is said to more than double Fable 5.2's pass rate at about one-sixth of the cost.

These are results claimed by the vendor, not independent measurements. According to OpenAI, on Cognition's FrontierCode 2.0, with both models at maximum reasoning effort, Sol came within half a point of Astra at about one-fifth of the cost.

Ultrafast and usage limits

Ultrafast mode speeds up token generation by up to 4.5 times, at the price of burning through the usage limit faster. The $200-a-month Pro plan gets it for the first time, but only for Sol, while the $500 plan gets it for both models. In Enterprise and Edu, a workspace admin has to enable the feature.

Safety caveats

According to SiliconANGLE, both models reached the Critical threshold for cybersecurity in OpenAI's Preparedness Framework, so tighter safeguards apply to higher-risk assistance. Those safeguards are also said to restrict some legitimate vulnerability research. OpenAI plans to bring the models to Daybreak, its program for vetted defenders.

Astra is the first externally released OpenAI model to reach the High threshold for AI self-improvement. Internally, the company has deployed a tool called Control Room that can halt model output. In two alignment tests, the new models were more likely than GPT-6 Astra to omit mentioning obstacles, such as a missing file. OpenAI says added instructions in Codex substantially reduce this, and it is still investigating the problem.

Context and what's next

The launch comes less than two months after OpenAI called off the October release of GPT-6.1 Astra, which in internal testing fell short on staying within the scope of a task and its authorization. Only GPT-6.1 Sol shipped then. For Polish companies, price matters most: a cheaper Astra changes the math for agents working on documents, though the per-task cost claims come from the vendor and are worth checking against your own data.

Share: