A plain prompt on a newer model beat many published coding tools.
Researchers tested 35 tools from this year's top software engineering conference against one automatically written prompt. The prompt ran on a newer model with no tuning and no repeated attempts. Depending on the measure, the single prompt won in 37 to 63 percent of the papers.
Why it mattersMuch of the scaffolding built around models is really a patch for last year's weakness. Before you build one, check whether the next model release simply erases the problem.
Read in the app — free, in 9 languages
Related stories
OpenAI's newest model finished Portal alone in about 24 hours.2026-09-08A quarter of DeepMind's cheating agent swarm turned whistleblower.
2026-09-08An AI-designed lung drug made patients' blood look years younger.
2026-09-07Opening a repository can run its code inside seven AI coding agents.
2026-09-07OpenAI's chief scientist says no lab should scale at full speed now.
2026-09-07