Gemini 4 Hydrogen matches Opus on CursorBench, but only at the costly setting

Google’s Gemini 4 Hydrogen matched Claude Opus 5.5 Max on CursorBench at 57.8 per cent, but the reported cost was $5.86 per task against Opus’s $13.43. The striking bit is the dial: Hydrogen’s score climbed sharply as its effort setting rose, and so did its bill.

Cursor Watch analysis

What happened

FourWeekMBA says it compared Cursor’s CursorBench leaderboard on 15 November. The benchmark uses ambiguous, multi-file coding tasks drawn from real Cursor sessions. At Hydrogen’s Low setting, it scored 27.9 per cent for $0.28 per task; at Extra High, it reached 57.8 per cent for $5.86. The article says that top score tied Claude Opus 5.5 Max, which cost $13.43 per task on the same leaderboard.

That is not a general model ranking. The numbers are results on Cursor’s task set, and Cursor cautions that small score differences may not be statistically meaningful. FourWeekMBA also notes that it did not test the models itself.

Why it matters

Hydrogen’s results show why a model name alone may tell developers little about the likely cost or capability of a coding run. In this comparison, turning effort up from Low to Extra High brought a 29.9-point score increase, but the cost per task rose to about 21 times as much, according to the article’s arithmetic.

The top setting still came in well below Opus on cost for a tied score. But at Low, Hydrogen’s much smaller bill came with a much lower score. The useful choice is not simply cheap model versus expensive model: it is which effort setting is good enough for the work in front of you.

Our read

This is a compelling comparison, not a universal verdict. CursorBench gives developers a concrete coding-task yardstick, while the large gap between Hydrogen’s effort settings is a reminder to test the setting as well as the model. Before changing a production workflow, compare representative tasks and the actual costs your own prompts produce. The bargain setting is only a bargain if it gets the job done.

What to watch

• Whether Cursor adds dated leaderboard snapshots, making changes easier to track.
• How Hydrogen performs on developers’ own coding tasks at different effort settings.
• Whether the score and cost gap between settings holds across other coding benchmarks.

Discussion spark: For coding agents, would you choose a much cheaper low-effort run or pay more for a higher score, and what result would justify the extra cost?

Independent WittyWires tracker for public updates about Cursor. Not affiliated with or endorsed by Cursor; this is not an official account.