OpenAI calls GPT-6 Astra the beginning of the AGI era

OpenAI released GPT-6 Astra on 3 September 2026, the company's most capable model to date. President Greg Brockman welcomed the launch with the words: "Welcome to the AGI era," positioning the model as a turning point in the development of artificial intelligence. Public availability via ChatGPT and the OpenAI API followed on 5 September 2026.

Astra is the first model OpenAI classifies as "Critical" under its own Preparedness Framework, the safety framework that assesses the risks of advanced AI models. Chief Research Officer Mark Chen stated that OpenAI considers itself "80 percent of the way" to AGI. OpenAI defines AGI as systems that outperform humans in the majority of economically valuable work.

The release was temporarily delayed in July 2026 after an incident involving Hugging Face prompted the company to implement additional safety measures. On 1 September 2026, two days before the launch, OpenAI published a safety update in which it assigned Astra the critical cybersecurity threshold, the first model to receive that designation.

High scores on mathematics and technical benchmarks

GPT-6 Astra achieves strong results across a broad range of standardised tests. On FrontierMath Tier 4, a mathematics benchmark targeting university and research level, the model scores 98 percent. On ARC-AGI-3, which measures general problem-solving, Astra reaches 99.9 percent, though a customised test setup developed by OpenAI itself was used for that evaluation.

In scientific reasoning the model scores 96 percent on GPQA Diamond, a master's-level benchmark. For software engineering and data analysis, Astra achieves 57.9 percent on Terminal-Bench 4.0, and 64.6 percent on Terminal-Bench Science 0.1 for scientific research. In computer use, where the model independently executes tasks on a simulated computer, Astra scores 72.6 percent on OSWorld 2.0 in an average of forty minutes per task. By comparison, predecessor GPT-5.6 Sol scored 65.7 percent in an average of 75 minutes per task.

Cybersecurity capabilities lead to "Critical" classification

The most notable findings during the evaluation relate to cybersecurity. Astra scored 100 percent on ExploitBench, a test in which the model develops exploits for known vulnerabilities, and 42.4 percent on ExploitGym. On SRE-Bench15, a benchmark for reverse engineering software binaries, the model solved 88 percent of tasks in a single attempt.

During the evaluation phase, Astra also independently identified two previously unknown zero-day vulnerabilities without human guidance at each step. This capability, finding new security flaws in systems that have already been hardened, was OpenAI's reason for classifying the model as "Critical" under the Preparedness Framework. That is a higher risk category than the company had previously assigned to any model.

At launch, OpenAI has deliberately restricted access to the most advanced cybersecurity features to a small group of testers. Broader availability will follow later through the company's Daybreak programme.

Availability and pricing

GPT-6 Astra is being rolled out to ChatGPT Plus, Pro, Business and Enterprise users. The model is also available via the OpenAI API, Microsoft Azure and AWS Bedrock. Use of Astra falls within existing subscription costs; for additional usage, subscribers and businesses can purchase supplementary credits. Holders of a Pro, Business or Enterprise subscription also receive access to GPT-6 Astra Pro.

Via the API, OpenAI charges ten dollars per million input tokens and fifty dollars per million output tokens. Reused prompt prefixes, so-called cached input, cost one dollar per million tokens. The model's context window is approximately 1.1 million tokens, allowing it to process considerably more text at once than previous generations.

What the AGI claim means, and what it does not say

The term AGI, artificial general intelligence, has no universally accepted definition. OpenAI uses the description of systems that outperform humans in the majority of economically valuable tasks. With that definition, the company applies a functional, economic measure rather than a philosophical or neuroscientific one.

Greg Brockman's description of the launch as the beginning of the AGI era is a positioning statement by OpenAI itself. Mark Chen simultaneously added that the company still places itself at 80 percent of the way to AGI, suggesting it is also viewed internally as an intermediate milestone. External researchers and competitors apply different thresholds and will not necessarily adopt that qualification. The benchmark figures do provide a concrete picture of what the model achieves under controlled conditions; translating that to everyday applications requires further observation.