OpenAI unveils GPT-6 Astra, touts advanced capabilities and higher pricing

OpenAI announced the launch of GPT-6 Astra, its latest flagship model, describing it as the most intelligent and aligned system it has produced. The company said the model achieved top scores on several demanding benchmarks, including a 98.6% result on the ARC-AGI-3 test and strong performance on coding, vision, and scientific tasks.

During a briefing, OpenAI President Greg Brockman suggested the model could be seen as marking the start of an artificial-general-intelligence era, though he stopped short of a formal claim. > "I think it’s not unreasonable to feel that we are now in the AGI era," Brockman said. He added that the term AGI is now treated as a conceptual goal rather than a contractual trigger.

Astra was trained on more than 100,000 GPUs at OpenAI’s Texas facility, the largest run the company has undertaken. The rollout will begin with enterprise customers in the Daybreak program and will expand to Plus, Pro, Business, and Enterprise users via the OpenAI API and AWS in the coming days. A Pro variant, Astra Pro, will also be offered, and API customers will have the option of zero-data-retention usage.

Pricing for the model is set at $10 per million input tokens and $50 per million output tokens, roughly 2.5 times OpenAI’s current promotional rates and comparable to Anthropic’s pricing for its latest model. OpenAI emphasized that the cost per task, rather than per token, is the key metric, noting that Astra often requires fewer tokens to complete a job.

Benchmark comparisons show Astra leading in several areas, such as a 74.1% score on the DeepSWE coding test and a 95.9% result on BenchCAD’s Vision2Code task, though competitors like Meta’s Muse Spark 1.3 posted a slightly higher coding score. Astra also demonstrated notable gains in scientific and research workflows, achieving 64.6% on the Terminal-Bench Science benchmark.

For developers, Astra introduces improvements to context handling, allowing the model to retain notes across windows and search earlier messages, reducing interruptions during long coding sessions. The model also supports concurrent questioning without halting ongoing work.

OpenAI said Astra performed better on internal alignment tests, avoiding unauthorized actions in impossible-task scenarios, but acknowledged that its written reasoning was harder to monitor than previous models. Chief Scientist Jakub Pachocki warned that advances in intelligence do not automatically translate to alignment progress and that scaling may be paused until monitoring confidence improves.

In cybersecurity evaluations, Astra surpassed earlier models in exploit generation tests, though the publicly available version will refuse certain advanced exploit tasks. Access to more permissive capabilities will be limited to vetted defenders through the Daybreak Blue program. OpenAI cautioned that users outside trusted programs may encounter slowdowns or blocks when attempting cybersecurity work. > "At launch, this is something that people should expect," Mia Glaese said.