Bigger AI models aren't getting much better at coding anymore

The AI industry has long operated on the assumption that more computing power and more training data equals a smarter model. For many general tasks, that held true. But in the specific domain of software engineering, that curve appears to be flattening — a pattern sometimes called 'diminishing returns.'

In practice, this means tools like GitHub Copilot, Cursor, or Claude may not feel meaningfully better with each new model version, even if the company behind them announces higher benchmark scores. Benchmarks measure performance on standardized tests, which don't always reflect the messy, complex coding problems developers actually face day-to-day. For solo developers and makers who rely heavily on AI coding assistants, this is a signal to think critically about which tool genuinely helps, rather than assuming the newest model is always the best choice.