GPT 5.5 sets new record in proofreading benchmark (revise.io)

GPT 5.5 has achieved a groundbreaking record in the ErrataBench proofreading benchmark, which evaluates the ability of large language models (LLMs) to identify and correct errors in human-written text. The benchmark assessed 64 model variants over 2,059 runs, measuring key performance metrics such as error detection and correction rates while incurring a total cost of $843 over a nearly seven-day runtime. This testing involved a robust dataset containing various writing errors across multiple domain-specific categories, ensuring a comprehensive evaluation of each model's proofreading capabilities. The significance of GPT 5.5's performance lies in its potential to enhance writing and editing processes across various sectors, making models not only faster but also more effective in identifying subtle errors. The benchmarking methodology allowed for an in-depth analysis of models' efficiency, comparing success rates with speed and cost-effectiveness. The findings indicate that LLMs can serve as valuable tools for proofreading, with the best variants displaying superior performance while maintaining cost-efficiency. By publicly sharing the results, including the associated code and dataset, the ErrataBench enables further experimentation and refinement, fostering ongoing innovation within the AI/ML community.