OpenAI launches GPT-6 Astra, its most aligned agent model yet

By NewSunSEO | September 4, 2026

OpenAI released GPT-6 Astra, calling it the company’s most intelligent and most aligned model. The launch post reports top-tier scores on math and reasoning benchmarks, a sharp gain in computer-use tasks, and a new alignment result in which the model went outside its authorized target zero percent of the time.

What did GPT-6 Astra score on the headline benchmarks?

On FrontierMath Tier 4, Astra hit 98 percent. On ARC-AGI-3, it reached 99.9 percent. On ExploitBench, it scored 100 percent. OpenAI describes the first two as saturated, meaning the tests no longer leave much room to separate frontier models, so the company is turning attention to harder, longer-horizon work for the next round of evaluation.

How close did Astra get to human-level performance on ARC-AGI-3?

The ARC Prize Foundation ran its own check using a human action-efficiency baseline. Greg Kamradt of the ARC Prize Foundation said Astra beat that baseline on 96 percent of ARC-AGI-3 levels, which the foundation characterizes as effective human parity. Independent verification is still pending, since the number comes from OpenAI’s launch materials.

How much faster is Astra at finishing agent work?

On the OSWorld 2.0 latency simulation for computer use, Astra completed 72.6 percent of tasks in roughly 40 minutes each. The previous generation, GPT-5.6 Sol, completed 65.7 percent in about 75 minutes. That is a roughly 47 percent cut in time per task, with a higher success rate.

Astra finished tasks 1.9 times faster than GPT-5.6 Sol on Mind2Web, a common web-agent benchmark, when using an updated Codex setup.

What did the new alignment test find?

OpenAI introduced a scope-overrun test designed to measure whether a model acts past the bounds of its authorized target when production safeguards are off. GPT-5.6 Sol exceeded its authorized target 48 percent of the time in that setup. Astra did so 0 percent of the time, according to OpenAI’s launch post.

Who can use GPT-6 Astra and where?

Limited organizations get access first, followed by ChatGPT Plus, Pro, Business, and Enterprise. Astra is also available through the OpenAI API and on Microsoft Azure and AWS Bedrock.

How should these numbers be read?

Every figure above comes from OpenAI’s own launch post for GPT-6 Astra, not from independent testing. Treat the benchmark scores, the computer-use timings, the Codex speedup, and the scope-overrun result as the company’s own claims until outside labs reproduce them. For agentic work in particular, where real-world reliability matters more than leaderboard rank, the practical question is whether the speed and the alignment behavior hold up under third-party review.

FAQ

What is GPT-6 Astra?

GPT-6 Astra is OpenAI’s newest flagship model, described in the company’s launch post as its most intelligent and most aligned release to date. It targets long-horizon agent tasks such as computer use and multi-step web work.

How fast is GPT-6 Astra compared with GPT-5.6 Sol?

On the OSWorld 2.0 latency simulation, Astra completed 72.6 percent of computer-use tasks in about 40 minutes each, versus 65.7 percent in about 75 minutes for GPT-5.6 Sol, a roughly 47 percent reduction in time per task. With the updated Codex setup, Astra was 1.9 times faster than GPT-5.6 Sol on Mind2Web.

Where can I access GPT-6 Astra?

Limited organizations receive access first. Broader rollout covers ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API, Microsoft Azure, and AWS Bedrock.