According to Ethan Mollick on X, citing the GDPval benchmark for GPT-5.4, the new model ties or beats human experts on professional tasks 82% of the time, as judged by independent experts, and can save an average of 4 hours 38 minutes on a 7-hour task after accounting for retries and one hour of human review (as reported by Ethan Mollick). According to Mollick, OpenAI did not update Figure 7 from GDPval for GPT-5.2 long-form task success, so he used GPT-5.2 Pro to extrapolate and update the chart showing operational time savings and expert-judged performance (according to Ethan Mollick). For businesses, this implies immediate ROI opportunities in knowledge work automation—delegating long-form tasks to GPT-5.4 with structured evaluation loops can compress cycle times, reduce expert billable hours, and expand throughput while maintaining expert-level quality on most tasks (as reported by Ethan Mollick).

According to Anthropic (@AnthropicAI), new research published on their Engineering Blog reveals that infrastructure configuration can significantly affect agentic coding evaluation results. The study demonstrates that variations in server environments and system settings can cause benchmark scores for agentic coding models to fluctuate by several percentage points, sometimes even exceeding the performance gap between leading AI models. This finding highlights the need for standardized infrastructure setups to ensure fair and reliable comparisons in coding model evaluations. As reported by Anthropic, these insights are crucial for organizations looking to accurately assess and deploy AI coding solutions.

According to God of Prompt on Twitter, recent findings by Scale AI published in May 2024 reveal that AI models are achieving over 95% accuracy on benchmark tests because many test questions are already present in their training data. This 'contamination' undermines the reliability of AI benchmark scores, making it unclear how intelligent these models truly are. As reported by God of Prompt, the industry faces significant challenges in evaluating real AI capabilities, highlighting an urgent need for improved benchmarking standards.