OpenAI launches GPT-6 Astra, claims major capability jump and declares "AGI era"

OpenAI released GPT-6 Astra and a higher-performance Astra Pro variant, rolling out first through its Daybreak program to select organisations, with wider access via ChatGPT Plus/Pro/Business/Enterprise and API/cloud platforms (AWS Bedrock, Azure) expected in the coming days. Standard pricing is $10 per million input tokens and $50 per million output tokens (double in fast mode), roughly 2.5 times the cost of GPT-5.6 Sol. OpenAI also classified Astra as "critical" under its Preparedness Framework for cybersecurity risk, the first model to hit that threshold, and added a cross-context-window note-taking feature to Codex.

Developers and enterprises now have concrete pricing and distribution details to evaluate for procurement, though a 2.5x cost increase means teams will need to weigh claimed performance gains against real budget impact. The "critical" cybersecurity classification is a genuinely new structural fact worth tracking, since it may affect how the model is deployed, audited, or regulated. Beyond that, the practical impact on most affected groups (consumers, researchers, competitors) is limited for now, since access is still gated and independent verification of the capability claims has not happened.

The "AGI era" framing from Greg Brockman is marketing language, not a technical claim that can be checked, and it is doing a lot of work to drive the headline. All performance benchmarks cited, including the 99.9% ARC-AGI-3 figure, are vendor-supplied under OpenAI's own self-defined test conditions with no independent replication, so treat the "substantially outperforms Anthropic" claims as promotional until third parties test them. No primary source links were captured in this extraction despite claims of official documentation, which limits how much of this can currently be independently verified.

1. 01Independent or third-party benchmark results (e.g. from research labs, LMSYS, or academic groups) comparing Astra against GPT-5.6 Sol and Anthropic's Fable 5 on the same tasks
2. 02Actual enterprise and developer uptake and public pricing feedback once API/Bedrock/Azure access opens broadly in the coming days
3. 03Details on what the 'critical' Preparedness Framework cybersecurity classification actually restricts or requires, and whether it triggers any regulatory or partner response