OpenAI this week unveiled GPT-6 Astra, a flagship model the company describes as its most intelligent and aligned system to date, and it released benchmark results showing dramatic gains on interactive reasoning tests.
OpenAI President Greg Brockman said the launch marks a new phase in capability and called it the start of an "AGI era," while leaving the definition of AGI to readers, as reported by OpenAI.
The company reported high marks on ARC-AGI-3, an interactive benchmark that measures agents' ability to explore, model, set goals and plan in novel, turn based environments.
Under the Standard harness, OpenAI reported Astra (max) at 62.7 percent on ARC-AGI-3 Semi Private for a reported cost of $26K, and it said a Provider Adapter harness that preserves opaque reasoning state raised observed scores to 99.9 percent for $19K, reflecting how system design affects measured results.
OpenAI also highlighted Astra's action efficiency, saying the model used fewer actions than the median human on 96 percent of ARC-AGI-3 levels and needed roughly 51.7 percent fewer actions per level on average.
Aidan Clark at OpenAI said Astra represented the company's largest training run to date, with pre training on more than 100,000 GPUs at the Stargate site in Texas, and that earlier models significantly supervised the process.
Implications Risks And Technical Details
OpenAI provided a wide set of benchmark comparisons. It reported Astra scored 74.1 percent on DeepSWE v1.1, 95.9 percent on BenchCAD Vision2Code with Python tools, and 64.6 percent on Terminal Bench Science, all presented alongside competitor figures.
The firm said Astra will roll out first to enterprise Daybreak users, then to Plus, Pro, Business and Enterprise tiers, and to the API and AWS in coming days, with API pricing set at ten dollars per million input tokens and fifty dollars per million output tokens.
OpenAI stressed trade offs in safety and monitoring. Chief Scientist Jakub Pachocki said written reasoning was harder to monitor in certain evasion tests, even as an internal check found Astra did not go outside authorized targets in impossible task scenarios, unlike the prior model.
The company also described cybersecurity findings. In controlled tests Astra produced higher exploit results and helped discover two previously unknown V8 vulnerabilities, though OpenAI said standard production access will refuse advanced exploit discovery and that Daybreak Blue provides vetted defensive access.
OpenAI highlighted harness differences. The Standard harness exposes visible notes the model chooses to keep, while the Provider Adapter and Responses API preserve opaque reasoning and use compaction for long contexts. In sandboxed PRO LONG runs Astra built custom parsers and small libraries to solve games, illustrating how external tools change agent performance.
