Skip to stories
America & the World Independent news
Technology

When is Gemini 4 Hydrogen’s extra thinking worth paying for?

Google added a fourth reasoning setting to its Gemini API. A higher benchmark score is one signal; the more useful measure for a software team is the cost per fix it can accept.

AI-generated illustration. Two identical breaks, with a longer path on one side, illustrate the extra effort a model can spend on the same bug. · United Herald / ImageGen
AI-generated illustration. Two identical breaks, with a longer path on one side, illustrate the extra effort a model can spend on the same bug. · United Herald / ImageGen

A software team testing Gemini 4 Hydrogen has a decision to make before it hands the model a bug: should it leave the reasoning setting at `high`, or ask the model to spend more effort? Google made Hydrogen generally available in the Gemini API on November 12. Its release notes list four settings: `low`, `medium`, the default `high`, and the new `xhigh`.

The setting matters even if the final patch is short. Under Google’s standard paid API rates, Hydrogen costs $1.50 per million uncached input tokens and $7.50 per million output tokens, including tokens used for internal reasoning. A hypothetical attempt that uses 20,000 uncached input tokens and 4,000 billable output tokens costs 6 cents in token charges. At the same input count, 8,000 billable output tokens would bring the charge to 9 cents. Those are examples, not measured token counts for either setting, and they exclude any separate tool charges.

A benchmark supplies a reason to try the new setting. Artificial Analysis reports a score of 54 for Hydrogen at `xhigh` on its Intelligence Index, a composite of nine evaluations. The score describes performance on that test mix. It cannot tell a team how often a patch will pass its own tests or how long a reviewer will spend checking it.

A useful trial would replay the same set of past bugs at both settings in an isolated copy of the repository. Give each attempt the same starting code, instructions, tools, and deadline. Before running it, decide what counts as a fix: the bug no longer reproduces, the existing tests still pass, and a reviewer would accept the change. Where possible, have the reviewer grade patches without knowing which setting produced them.

Then divide each setting’s total token charges, including those for failed attempts, by the number of fixes accepted. Log elapsed time and human review time separately. For example, if 100 attempts at 6 cents each yielded 50 accepted fixes, the token cost would be 12 cents per accepted fix. At 9 cents per attempt, the same 100 bugs would need to yield 75 accepted fixes to match that cost. Neither rate is a forecast. A team might still pay more for difficult bugs if it saves enough reviewer time, but the trial should show where that trade pays off.

This post is available in the following languages:

Link