> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI ships GPT-6 Astra as its president declares the AGI era begun
- URL: https://www.metatalks.ai/openai-ships-gpt-6-astra-as-its-president-declares-the-agi-era-begun/
- Published: 2026-09-07T08:53:00.000Z
- Updated: 2026-09-07T08:52:59.000Z
- Author: Al
- Tags: News, Frontier Models, AI safety, #newswire

**Independent scorers put Astra level with GPT-5.6 Sol, and its near-perfect result on the benchmark named for AGI came only through a purpose-built harness.**

OpenAI released GPT-6 Astra last week as its new top-tier model, and company president Greg Brockman used the launch briefing to attach the AGI label to it. He told journalists it was not unreasonable to think the AGI era had already begun, and that in a couple of years people might date AGI to roughly this moment and roughly this model.

The model went first to a small group of organizations, with OpenAI saying at launch that its ChatGPT subscription tiers, its own API and AWS would follow within days, at the same API price as Claude Fable: $10 per million input tokens and $50 per million output.

Astra is also the first model OpenAI has placed above its own [critical cybersecurity capability threshold](https://www.theverge.com/ai-artificial-intelligence/989601/openai-gpt-6-astra-release?ref=metatalks.ai), the point at which the company says a model can find and exploit vulnerabilities in well-defended systems without human guidance. Access at that level goes first to a set of trusted defenders.

OpenAI called Astra the most intelligent and most aligned model it has built, pitching it on multistep agentic work, real codebases and computer use, and said it had already helped settle mathematical questions that had gone unanswered for years. Mathematicians posting their own results gave one concrete case: the bound on prime gaps, the distance within which infinitely many pairs of primes are proved to fall, came down from 246 to 186, with the argument checked in Lean.

Artificial Analysis put Astra at 61 on its [Intelligence Index](https://simonwillison.net/2026/Sep/3/gpt6-astra/?ref=metatalks.ai), five points short of Claude Fable 5.1 in its maximum setting with fallback and behind the maximum configuration of Meta's Muse Spark 1.3\. On the same firm's coding-agent index it does better, scoring two points above Sol at about the same cost per task.

Evaluators at ARC Prize scored the model 62.7% on ARC-AGI-3 in the benchmark's default harness, and 99.9% when it ran through OpenAI's own provider-adapter harness, which keeps the model's reasoning state between requests. On 96% of the levels it completed, they said, it needed fewer actions than the median human.

Gary Marcus, who has argued for [neurosymbolic world models](https://garymarcus.substack.com/p/hot-take-on-gpt-6-astra) for nearly a decade, called Astra impressive and said multiple reports pointed to a genuine advance. He also holds that success on ARC-AGI, whatever the benchmark's name, is not proof of AGI, and expects the model's best work to stay in domains where an answer can be checked.

The system card describes Astra as better aligned and, at the same time, less monitorable in its chain of thought. OpenAI's chief scientist Jakub Pachocki told the briefing that progress in intelligence does not guarantee progress in alignment, and that monitoring these systems is getting harder.