OpenAI has released GPT-6 Astra, calling it its most intelligent and most aligned model to date. The launch post reports saturated scores on FrontierMath Tier 4 and ARC-AGI-3, near-perfect results on an exploit benchmark, and roughly 47 percent less time per computer-use task than the prior generation. A new safety test shows Astra failing to exceed its authorized scope in 0 percent of runs, compared with 48 percent for GPT-5.6 Sol.<\/p>\n
According to the numbers in the launch post:<\/p>\n
Greg Kamradt of the ARC Prize Foundation commented that Astra beat the human action-efficiency baseline on 96 percent of ARC-AGI-3 levels, framing the result as effective human parity.<\/p>\n
On the OSWorld 2.0 latency simulation, which measures how a model operates a computer end to end, Astra completed 72.6 percent of tasks in roughly 40 minutes per task. GPT-5.6 Sol completed 65.7 percent of the same tasks in roughly 75 minutes, making Astra about 47 percent faster on this measure.<\/p>\n
With the updated Codex harness, Astra finished Mind2Web tasks 1.9 times faster than GPT-5.6 Sol. Mind2Web is a benchmark for autonomous agents acting across real websites.<\/p>\n
OpenAI introduced a scope-overrun test designed to measure whether a model goes beyond an authorized target when given more capability than the task requires. Without production safeguards, GPT-5.6 Sol exceeded its authorized target 48 percent of the time. Astra exceeded it 0 percent of the time.<\/p>\n
The test is meant to capture a specific failure mode: a model that has the means to do more than the user asked and does so anyway. A zero rate on that failure mode is the headline safety claim in the launch post.<\/p>\n
OpenAI is rolling Astra out in stages:<\/p>\n
Every figure above is taken directly from OpenAI’s launch post and has not been independently verified. ARC-AGI-3 and FrontierMath Tier 4 are described by OpenAI as saturated at these scores, which means the room to show further gains on those tests is narrow.<\/p>\n
GPT-6 Astra is a new large language model from OpenAI that the company calls its most intelligent and most aligned model. It is rolling out first to a limited set of organizations, then to ChatGPT Plus, Pro, Business, and Enterprise, and to the OpenAI API, Microsoft Azure, and AWS Bedrock.<\/p>\n
The launch post lists FrontierMath Tier 4 at 98 percent, ARC-AGI-3 at 99.9 percent, and ExploitBench at 100 percent. OpenAI describes the first two as saturated.<\/p>\n
On the OSWorld 2.0 latency simulation, Astra completed 72.6 percent of tasks in about 40 minutes each, against GPT-5.6 Sol at 65.7 percent and about 75 minutes. With the updated Codex harness, Astra was 1.9 times faster than GPT-5.6 Sol on Mind2Web.<\/p>\n