OpenAI Releases GPT-6 Astra Model, Suggests It Could Be AGI

OpenAI on Thursday started releasing its new GPT-6 Astra model, billing it as a major advance in commercial tasks from financial modeling to engineering design.

The new flagship model, which OpenAI said also has significant cybersecurity capabilities, will initially be available to a limited number of organizations, including those that are part of OpenAI’s Daybreak Access cybersecurity program, the company said in a blog post. Astra will be made available in the coming days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI application programming interface and Amazon Web Services, the company said.

OpenAI said that the new model showed significant improvements in areas like using a computer, software programming and professional work like creating presentations and spreadsheets. The company added that the model performs at the highest level on benchmarks including FrontierMath Tier 1, which tests mathematics ability, ARC-AGI 3 for reasoning ability and ExploitBench for cybersecurity.

In a press briefing, OpenAI cofounder and President Greg Brockman said that the question of when the industry achieves artificial general intelligence, or AI that’s as good as humans at most tasks, has been a fuzzy one. However, if we “fast forward a couple years and look back,” he said, the milestone of achieving AGI “might be around this time and about this model.”

The Information previously reported that GPT-6 Astra uses a new technique that improves its performance but could also mean that the model would reveal less of its “thinking,” meaning that it could be more difficult to monitor for signs of bad behavior.

OpenAI said in its blog post Thursday that its “evaluations found Astra’s written reasoning harder to monitor than GPT-5.6 Sol’s.” OpenAI said that it attributed this to “Astra’s greater control over written reasoning on simpler tasks and ability to solve problems with fewer written steps.” However, it added that GPT-6 Astra wasn’t as good at hiding its thinking process on more complex tasks.

During the press briefing, OpenAI chief scientist Jakub Pachocki wouldn’t comment on whether this change was due to the new technique The Information had reported on. However, he said that over time it would naturally get more difficult to monitor models’ thought processes, either because the models would become more aware that this was happening and hide their thinking processes or because they would be able to perform harder tasks “using fewer language tokens or all language tokens.”

In response to a question during the briefing, Brockman said that GPT-6 Astra had gone through the government’s recently established voluntary testing framework, which provides a system for top AI labs to share their models with the government before releasing them to the public and partners.

GPT-6 Astra will be 2.5 times more expensive than OpenAI’s previous flagship model, GPT-5.6 Sol, at $10 per million input tokens and $50 per million output tokens. That’s the same pricing as Anthropic’s recently-released Fable 5.1 model.

However, Brockman hinted that OpenAI and the industry as a whole could eventually move away from per-token pricing for AI models.

“Pricing tokens doesn’t make any sense,” he said. “Our tokens are not the same as our competitors’ tokens, they’re not the same between different model families.” He added, “What you actually want is…the price per task is what matters.”

Brockman ended the briefing on an enthusiastic note, telling journalists, “welcome to the AGI era!”