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.
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
- Google DeepMind: Gemini 4 Hydrogen launch
- Google Search: Hydrogen rollout
- Google AI Studio: Hydrogen model specification
- Google AI Studio: Gemini API pricing and caching
- Google AI Studio: Output and thinking token fields · Thinking token pricing
- Google DeepMind: Hydrogen model card
- Google Antigravity: Hydrogen and Orchestrator

Top comments (0)