Cursor Adds Gemini 4 Hydrogen, Hitting 54.3% at One-Tenth Opus Cost

Cursor has added Gemini 4 Hydrogen, Google’s new Flash-tier coding model, to its lineup. At the default high reasoning setting, Hydrogen scores 54.3% on CursorBench 4.0, up 14.7 percentage points from Gemini 3.8 Flash. The gain gives developers a lower-priced option for agent runs, parallel workers, and large-repository tasks.

Flash pricing reaches frontier scores

Cursor lists Hydrogen at $1.50 input, $7.50 output per million tokens. Gemini 4 Argon costs $2 and $10, while Anthropic’s Opus 5.5 costs $15 and $75. Hydrogen therefore costs 25% less per token than Argon and one-tenth as much as Opus.

At those rates, a run using 100,000 input tokens and 20,000 output tokens would cost about $0.30 with Hydrogen, $0.40 with Argon, and $3 with Opus. Actual bills also reflect reasoning depth, cache behavior, retries, and Cursor plan terms.

Inside the 54.3% score

CursorBench 4.0 evaluates ambiguous, multi-file tasks drawn from real Cursor sessions. Its cases exercise planning, tool calls, debugging, code editing, and repository navigation inside Cursor’s agent harness. The score therefore reflects integrated agent performance across a workflow.

Hydrogen’s score sits 1.5 points above Argon at medium effort, 3.5 points below Opus at max, and 6.1 points below Carbon at high. Since the effort settings differ, a controlled ranking would require matched configurations.

Google reports a similar pattern on the Artificial Analysis Intelligence Index. Hydrogen scores 51 at its default high setting, compared with Argon’s 50, and reaches 54 at xhigh. That index covers broader model capabilities and provides a separate performance signal.

Four reasoning budgets

Hydrogen exposes low, medium, high, and xhigh reasoning levels, with Cursor selecting high by default. Lower settings reduce latency and token use for routine work, while xhigh allocates more computation to difficult planning and debugging. Higher effort can produce more billed tokens even when the published per-token rate stays fixed.

Google advertises native support for 1 million input tokens and up to 1 million output tokens. That capacity can accommodate large codebases, long tool traces, and extensive generated changes. The usable limit in practice depends on the context and output allowances Cursor exposes in each mode.

Built for fan-out

Google positions Hydrogen for high-volume worker and verifier roles in multi-agent systems. A coordinator such as Carbon can divide a project into tasks while several Hydrogen agents edit, test, and review changes in isolated workspaces.

• Routine implementation: Handle small features, targeted fixes, and predictable refactors with lower latency.
• Parallel execution: Assign independent modules or subtasks to multiple workers while a stronger model manages the plan.
• Verification: Review diffs, run tests, inspect failures, and report issues at lower token cost.
• Large-context work: Read broad sections of a repository without immediately moving the task to a more expensive model.

Speed compounds the discount

Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at the default high setting. Launch partners report speed gains ranging from 2 to 2.8 times, alongside token reductions of 12% to 23%. These vendor and partner measurements depend on repository size, tool latency, task mix, and reasoning level.

On workloads with comparable success rates, higher throughput and lower token use can reduce the cost per completed task beyond Hydrogen’s 25% per-token discount against Argon.

Measure it on your repository

CursorBench is a first-party evaluation built around Cursor’s tasks and tooling. That alignment makes it relevant to Cursor users, while its prompts, tool implementations, scoring, and mixed effort settings limit broader head-to-head conclusions.

Teams evaluating Hydrogen can run a representative set of repository tasks and track completion rate, latency, input and output tokens, retries, test results, and reviewer corrections. Its published profile makes it a practical default candidate for routine Cursor Agent and chat work, with Carbon or Opus reserved for tasks that demand deeper planning.