article --summary
The summary
The author tests OpenAI's new GPT-6.2 Sol against Anthropic's Claude Opus 5.5 on three structured tasks: a month-end billing close, a billing library repair, and a messy lab data analysis. Opus 5.5 scores higher on the agent task, while the two tie on the coding and data work, so accuracy differences are small.
The real gap is cost and speed. Sol was dramatically cheaper per run and faster in every test, which the author says matches OpenAI's price claim better than its accuracy claim. The main caveat is that Sol missed a customer credit in the billing close, a mistake that could matter in real work.
The author's advice is to start with Sol for work where errors are easy to catch, run tasks twice to compare, and keep Opus 5.5 as the final check where mistakes would cause serious downstream problems.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.