Cursor adds Gemini 4 Hydrogen: 54.3% on CursorBench

Cursor has added Google's Gemini 4 Hydrogen to its AI code editor and reports that, at its default effort, the model scores 54.3% on CursorBench, up from 39.6% for Gemini 3.8 Flash, the earlier model Google presents Hydrogen as an upgrade over (announcement). CursorBench is Cursor's internal evaluation, which its results page describes as ambiguous, multi-file tasks taken from real Cursor sessions. Effort here means how much the model is allowed to reason before acting, and Google's launch post sets Hydrogen's default at high, one of four levels.

Cursor's results page shows why the effort level has to travel with the score. Hydrogen scores 57.8% at its highest setting, 54.3% at high, 46.7% at medium and 27.9% at low, with average cost per task falling from $5.86 to $0.28 across that range. Gemini 4 Carbon, one of Google's frontier models, currently leads the table at 60.4% for $9.22 per task. On Cursor's own numbers, Hydrogen at high recovers most of Carbon's score for roughly 60% of its cost, while dropping to medium cuts the cost by about two thirds and gives up nearly eight points. Cursor notes that small differences between scores may not be statistically meaningful, and these are Cursor's measurements on its own benchmark, read from its results page on November 14.