
Gemini 4 Hydrogen: Google's fast agent model, its benchmarks and its price
Gemini 4 Hydrogen from Google: $1.50 and $7.50 per million tokens, a 1M context, four thinking levels and agent benchmarks. Data as of November 15, 2026.
gemini-4-hydrogen) is Google’s model for everyday agent and coding work, published on November 12, 2026. It is the fast-execution model in the Gemini 4 family, with a lower per-token price than Gemini 4 Carbon on Google’s pricing page, and Google calls it its “daily driver” for agentic workflows.The data on this page is as of November 15, 2026, and it changes often. Confirm prices and limits on the pages listed under Sources.
What is Gemini 4 Hydrogen and when did it come out?
Google introduced it on November 12, 2026 (the page was updated on the 13th) with the line “bringing cutting-edge intelligence to everyone”. According to the announcement, Gemini 4 Carbon and Gemini 4 Argon handle deep, long-horizon reasoning, while Hydrogen is built for fast execution in agentic workflows.
Google compares it mostly with Gemini 3.8 Flash, an earlier model. To understand what choosing a fast model over a deep one means, read small vs. large models.
What is it for?
- Everyday agents: multi-step tasks like researching, analyzing data or navigating a repository, where speed matters. To understand how they chain together, see reasoning models.
- Code: Google positions it as its everyday model for programming.
- Knowledge work with visual files: Google reports results on a benchmark of complex visual files (see benchmarks).
- Spreadsheets: in Gemini in Google Sheets, Google says outputs are generated over 30% faster than with previous models.
- Low-wait responses: Google says that at its default thinking level it finishes tasks 2.5 times faster than Gemini 3.8 Flash. Latency shapes the experience, as LLM latency explains.
If you want Google’s assistant for your daily work, see Gemini for designers; this page covers the model.
What is the quick spec sheet?
| Item | Value (as of Nov 15, 2026) |
|---|---|
| API ID | gemini-4-hydrogen |
| Lab | |
| Date | November 12, 2026 |
| Context | 1 million input tokens; the announcement mentions up to 1 million output |
| Max output | “64K / 1M” on the models page, with no explanation of each value |
| Thinking levels | low, medium, high (default), xhigh |
| Standard price | $1.50 input / $7.50 output per million tokens |
| Open or closed | Closed (API and Google apps) |
What benchmarks did it publish and how do you read them?
These scores are reported by Google in its announcement, and some are measured by a third party. Each lab’s benchmarks are not comparable with each other, so do not build a cross-lab ranking from them. To learn how to read them, see how to read AI benchmarks.
| Benchmark | Hydrogen score | Who measured it |
|---|---|---|
| Artificial Analysis Intelligence Index | 51 at high thinking; 54 at xhigh | Artificial Analysis (linked from the announcement); reported by Google |
| Terminal-Bench 4.1 | 57% (Gemini 3.8 Flash: 18.2%) | Reported by Google |
| HLE-Diamond | 58.8% | Reported by Google |
| Agents’ Last Exam v1.1 | 71.6% | Reported by Google |
| Multimodal Knowledge Agent (MKA) | 53.4% | Reported by Google |
| Vals Tax Agent Bench | 52.3% (rank 1 overall, per Google) | Vals; reported by Google |
How to read them, using what Google says in the announcement:
- Agents’ Last Exam v1.1: Google gives the score “for end-to-end professional workflows” and does not describe the test.
- Multimodal Knowledge Agent: Google presents it as a new benchmark of agentic knowledge work over complex visual files.
- Terminal-Bench 4.1: the announcement gives the score without saying what it measures. Look at the comparison: 57% versus 18.2% is against an earlier model from Google itself, not against other labs.
- Artificial Analysis Intelligence Index is a composite index from a third party. Google says Hydrogen beats Argon at high thinking (51 versus 50).
- For HLE-Diamond and Vals Tax Agent Bench the announcement gives the score without a short definition; treat them as data reported by Google.
Google also cites a 94% honesty rate on an internal agent evaluation and says false refusals on HealthBench Pro stay under 0.4%. Those are the lab’s own evaluations.
How much does it cost and how much context does it have?
According to Google AI Studio’s pricing page, as of November 15, 2026, per million tokens (output includes thinking tokens):
| Tier | Input | Output | Context caching |
|---|---|---|---|
| Standard | 1.50 | 7.50 | 0.075 |
| Batch or Flex | 0.75 | 3.75 | 0.0375 |
| Priority | 2.70 | 13.50 | 0.135 |
Points to keep in mind:
- Flat price: the announcement says the price is the same across the whole 1-million-token window, with no surcharge for long prompts. That contrasts with models that charge more past a threshold.
- Free tier: the pricing page says input, output and caching are free of charge on Standard and Priority in the free tier, and that Batch and Flex are not available there.
- No date-based changes: unlike Carbon, whose price goes up on December 1, 2026 according to the pricing page, the page lists no scheduled change for Hydrogen.
- Context caching is implicit.
To understand how the bill is built, read cost per token, tokens and context window.
When should you pick it and when not?
Pick it when:
- You need agents or code with a good balance between speed and cost.
- You process a lot of text or files in a single request and want a flat price.
- You want to test without paying on the API’s free tier.
- You measure user experience and every second of waiting counts.
Skip it when:
- Your task is deep, long-horizon reasoning: Google presents Gemini 4 Carbon and Argon for that.
- You need open weights: see open vs. closed models.
- Your task is very simple and enormous in volume: a smaller model can be cheaper there.
Example
Illustrative example: a product designer wants an agent to review 30 screens of a prototype and list text and component inconsistencies. She runs the same task on Hydrogen at the low, high and xhigh levels. For each run she notes how long it took, how many tokens it used and how many real inconsistencies it found, checking against her own review.
What to try tomorrow: build a table with those three runs and decide with your own numbers, not the announcement’s, which thinking level to keep as the default in your flow. With a flat price of $1.50 and $7.50 per million tokens, each run’s cost is easy to calculate.
Common mistakes
- Using xhigh every time without measuring: a higher level usually means more tokens and more time.
- Reading “57% versus 18.2%” as a comparison with other labs: it is against Gemini 3.8 Flash.
- Comparing its scores with another lab’s as if they were the same test.
- Thinking “daily driver” means “the best”: it is Google’s label for the everyday model.
- Assuming the free tier is fit for production: the page offers it at no cost, but check limits and terms before relying on it.
- Not rechecking prices: they can change.