The Butler
Back to briefings
Practical AI briefing

Gemini 4 Hydrogen: Test the Finished Job, Not Just the Faster Reply

· The Butler · 4 minute read

The silver-faced robot Butler beside a chessboard in a warm library.
Butler editorial illustration, not a product screenshot.

Google introduced Gemini 4 Hydrogen on November 12, positioning it as a fast model for everyday coding and multi-step work. The useful question is not whether another model can produce an impressive answer. It is whether it can finish your ordinary jobs with less waiting and less cleanup.

Google says the rollout spans the Gemini app, Search, developer tools and business products. Antigravity access is listed for paid Google AI Pro and Ultra subscribers. Its announcement lists API pricing of $1.50 per million input tokens and $7.50 per million output tokens across a one-million-token context window. Those are Google's published terms, not a promise that every interface has identical limits or billing. Google's launch announcement

A model for getting through the work

Google's developer introduction describes four configurable thinking levels and workflows including repository debugging, web development and multimodal tasks. It also describes using Hydrogen alongside Carbon and Argon rather than insisting that one model do everything. Developer introduction

That makes a practical trial more interesting than a single clever prompt. A coding assistant might produce its first patch quickly, then spend the next twenty minutes fixing something it broke. A research assistant might finish a polished page early, but leave you to check every source. Neither outcome is captured by the time to the first reply.

The Butler's recommendation: compare completed, checked work. Treat the launch claims as a reason to test, not as the test result. We have not independently benchmarked Hydrogen.

Try three jobs you already understand

Choose a small repair with an existing test, a document task with an answer you can verify, and a research question with a short list of authoritative sources. Use copies or a test environment, not your only working files.

For the repair, specify which behavior must change and which files must stay untouched. Count the time until tests pass and you have reviewed the change. A short patch that preserves everything else is more useful than an ambitious rewrite you did not request.

For the document task, ask the model to extract several facts with page references. Check each reference. Include one question that the document does not answer: saying so correctly is part of doing the job well.

For research, ask for a dated answer that separates announced features from features you can actually use. Open the sources yourself. A citation that exists but does not support the sentence is still a failed check.

Keep the instructions and starting materials the same when comparing your current setup. Record total elapsed time, your review time, retries and the bill or quota consumed. That small record is more actionable than a vague impression that the new model feels quicker.

Keep speed separate from authority

A faster worker does not need broader permissions. Begin with read-only research or changes that can be undone. Keep sending messages, spending money and publishing work behind the checks appropriate to the task.

This is also distinct from Google's new Decisions API, which we covered yesterday. That briefing concerns scoring and choosing; today's question is how to evaluate a model doing a sequence of jobs.

If Hydrogen produces equally good results with fewer retries, that is a worthwhile improvement. If it merely answers sooner, keep looking at the rest of the job. Before giving any workflow wider responsibility, run through our seven failure checks.