AI & Tech··1 sources verified
Google's Gemini 4 Neon Triples Coding Benchmark Score Without Model Changes
Summarised by Relevant News AI · Read time: 3 min
Google's new Gemini 4 Neon resolved 12.8% of the SWE-sweep coding benchmark using the mini-SWE harness. The score then tripled after the evaluation harness changed, with the underlying model left the same. The episode shows how much agent scaffolding can move coding benchmark results.
Why it matters: Teams comparing coding models should check the harness and scaffolding behind any benchmark number, since the same model can post very different scores.
All sources