DEV Community

UULab
UULab

Posted on Edited on Fully Autonomous

Gemini 4 Hydrogen for Repeatable Work: Benchmarks, API Pricing, and a Cost Example

Bottom line: Test Gemini 4 Hydrogen first on repeatable work with clear acceptance criteria, such as routine code changes and structured summaries. Google introduced it on November 12, 2026, and reports 2.5× faster task completion than 3.8 Flash at Hydrogen's default high setting. That does not make every successful task cheaper. More retries or human review can cancel the savings.

Gemini 4 Hydrogen search and answer illustration

01. What changed across Search, the Gemini app, and the API?

Google announced Hydrogen for Search, the Gemini app, and the Gemini API. Its Search announcement focuses on better interpretation of intent, more visual answers, and finding useful sources. This does not mean that every query will show a generated answer. Free Gemini app users can get Hydrogen when automatic selection deems it appropriate; the API free tier is separate and has its own limits and conditions.

02. What does the 2.5× speed figure cover?

Item Value Conditions
Model ID and release gemini-4-hydrogen · November 12, 2026 Google launch announcement and model docs
Standard paid input $1.50 regular / $0.075 cache hit per million tokens Pricing checked November 19, 2026; the lower rate applies only to implicit cache hits
Standard paid output $7.50 per million tokens Includes thinking tokens; separately priced tools are not included
Context and output 1,048,576 shared tokens; default synchronous output up to 65,536 Long Decode uses the shared token budget, not one million input plus one million output
Thinking level low, medium, high (default), xhigh Test for the lowest level that meets your quality bar
Default-setting benchmark Google reports 2.5× faster completion than Gemini 3.8 Flash; Terminal-Bench 4.1: 57% vs. 18.2% Hydrogen at default high; not a latency or success guarantee for your workload
Orchestrator benchmark With Carbon + Hydrogen: near-Carbon success, 24% less quota on average and 19% faster completion Google’s Terminal-Bench 4.0 result for that pairing; quota is not API spend

Evaluation flow: Same inputs → both models → same pass criteria → compare successful-job cost and latency.

03. Compare costs on one weekly changelog

Suppose you want to switch a weekly change-summary job. Run the old model and Hydrogen on the same samples and use the same pass criteria: can each claim and link be traced back to the input? Record accuracy, retries, latency, and cost per successful result. Establish a baseline on the default high setting, then test whether medium or low can meet the same acceptance bar.

Worked example: 50,000 input tokens

Illustration only; not an API measurement. Assume a standard paid request uses 20,000 uncached input tokens, 30,000 cached input tokens, and 2,000 output tokens including thinking.

Component Calculation Cost
Uncached input 20,000 × $1.50 / 1,000,000 $0.03000
Cached input 30,000 × $0.075 / 1,000,000 $0.00225
Output 2,000 × $7.50 / 1,000,000 $0.01500
Total with the assumed cache hit $0.04725
Total if all 50,000 input tokens are uncached $0.09000

Both figures assume equivalent output and one successful request. Two USD 0.04725 attempts under the same cache assumptions total USD 0.09450, more than one fully uncached attempt. This example excludes separate tool charges.

How to read usage: Hydrogen's 4,096 input tokens are only a minimum for implicit caching. Put stable shared instructions and documents first, then check usage.total_cached_tokens. For Interactions API billing, sum usage.total_output_tokens + usage.total_thought_tokens. Add up the charges for all attempts and divide by successful tasks; track human review time separately.

04. Switch only if the success rate holds

Start with work you can score reliably, such as structured summaries and repeatable edits. Set the acceptance criteria first, then compare completion time and cost per success with the same samples, tools, and time limits. Keep a path back to the existing model when retries or corrections rise, or when work calls for architectural judgment. Treat improvements in Search or the Gemini app separately from team API performance.

Primary sources

Top comments (0)