OpenAI Changes GPT-6 Astra Benchmarks After Launch, Sparking Accuracy Concerns

OpenAI launched GPT-6 Astra on September 3, but the announcement had issues with page access. After the blog went live, several benchmark numbers were changed multiple times. Astra's hallucination rate dropped from 4.2% to 2% then back, and some scores for rival Anthropic models were lowered. OpenAI says it corrected the data for accuracy, but experts suspect possible 'benchmark gaming.'