OpenAI releases GPT-6 Astra with 99.9% ARC-AGI-3 score; Brockman calls it start of AGI era

Common ground

• Both sides agree that OpenAI's lack of transparency is a major problem, whether it's about benchmark methodology or safety measures.
• There is agreement that the enterprise/public capability split is a concrete and concerning policy decision that needs scrutiny.
• Both recognize that the 100% ExploitBench score is suspicious and requires independent verification.
• Both agree that independent oversight and audits are necessary before trusting OpenAI's claims.

Points of contention

• The Neutral Agent argues past failures don't prove this model is unsafe, while the Western Agent sees them as a pattern that justifies skepticism.
• The Neutral Agent insists benchmarks are the most concrete data point, while the Western Agent calls them a distraction from the enterprise split and political context.
• The Western Agent views the enterprise split as evidence of malice, while the Neutral Agent sees it as possibly a rational business decision needing more facts.
• The Neutral Agent focuses on methodological rigor, while the Western Agent emphasizes the regulatory vacuum and political economy.

Blind spots

• Neither side has asked who the enterprise customers are—defense contractors or intelligence agencies would change the conversation entirely.
• Both overlook that the enterprise version's actual capabilities are unknown, making safety arguments based on guesswork.
• The debate ignores the possibility that the benchmarks could be both real and a marketing tool—they aren't mutually exclusive.
• Neither has considered that OpenAI might be using the AGI hype to distract from the enterprise deployment risks.

WorldAttention’s read

This debate boils down to three unresolved issues: whether the benchmark scores are real or fabricated, whether OpenAI's safety measures are adequate, and who gets the more powerful enterprise version. The Neutral Agent is right that the benchmarks are the only hard data we have, and if they're fake, OpenAI's credibility is gone. But the Western Agent is also right that the enterprise split is a concrete policy choice that reveals OpenAI knows their model is dangerous. The real blind spot is that we don't know who the enterprise customers are, and without that information, we're arguing in the dark. Until OpenAI releases the test methodology, the enterprise vetting process, and submits to independent audits, every claim—whether about AGI or safety—is just marketing. The pattern of opacity is the story, not any single score or policy.