Verdict On AGI: A Critical Audit of OpenAI's Claim to Have Achieved Artificial General Intelligence

Verdict on AGI is a 47-page critical audit of OpenAI's claim — anchored in the August 2026 TIME interview with Sam Altman and the September 3, 2026 release of GPT-6 Astra — that the company has achieved or is about to achieve Artificial General Intelligence. The paper applies a strict four-tier verdict scale (HONORED / PARTIALLY HONORED / EXPOSED AS MARKETING / INVALID) to eight major OpenAI AGI-anchored claims from 2018 through September 2026, using Definition D (flexible, sample-efficient generalization to genuinely novel environments) as the primary bar. The central finding: across all eight claims, zero meet the strict bar, two are partially honored, five are exposed as marketing, and one is invalid; Astra is not an exception but the eighth instance of a recurring pattern with five structural defects — no operational definition, selected benchmarks, closed-door demos, executive estimates without methodology, and internal designations coinciding with commercial pressure.

The strongest evidence-anchored AGI claim is not OpenAI's — it is Anthropic's, which as of September 5, 2026 holds the #1 spot on BenchAlign (Claude Fable 5.1) and leads SWE-bench Verified through Claude Opus 4.7 (87.6%), backed by the only public binding safety commitment (the Responsible Scaling Policy with ASL-3 deployment-pause clause) and the September 1, 2026 Enterprise Frontier Safeguards. The paper's secondary contribution is a current-as-of-September-5-2026 model lineup covering OpenAI's GPT-6 Astra and GPT-5.6 Sol, Anthropic's Claude Fable 5.1 and Opus 5, Google DeepMind's Gemini 3.8 Flash, xAI's Grok 4.6, Meta's Muse Spark 1.3, DeepSeek's V4-Pro 0813, Alibaba's Qwen3.8-Max, and Mistral's Medium 3.5. The pattern itself is the finding: when the most consequential capability claim in 21st-century technology is examined against its own published evidence, AGI is what OpenAI says it is, and what OpenAI says it is keeps moving.