Google shipped a new AI model on Nov. 12, 20261. Koray Kavukcuoglu is Google DeepMind's chief AI architect. He built Gemini 4 Hydrogen for one job: fast coding and agent work, the kind developers run dozens of times a day. The price surprised many of them.
Hydrogen charges $1.50 for every million input tokens. Output tokens cost $7.50 per million1. That price holds flat across its full 1 million-token window.
A token is the chunk of text a model reads or writes, close to three-quarters of a word. Anthropic charges a fifteenth of that for Claude Haiku 5.5, a rival built for the same routine work2.
Developers Meet a Steeper Price
Google says the extra money buys speed. Hydrogen finishes coding tasks two and a half times faster than Gemini 3.8 Flash, the model it replaces1. At its default setting, Hydrogen solved 57 percent of tasks on Terminal-Bench 4.1, a test of real coding and terminal work. Flash solved just 18.2 percent.
Claude Haiku 5.5 costs less. It runs $0.10 per million input tokens for a short prompt2. Output costs $0.50 per million.
That price holds under 100,000 tokens of prompt. Anthropic checked those numbers on its own pricing page, Nov. 14, 2026. Above that length, Haiku's price jumps to $0.50 and $2.50.
Hydrogen keeps one flat price across its whole window, whatever the prompt's length1.
Hydrogen Charges 15 Times Haiku's Price for the Same Million Tokens
List price per million tokens, Gemini 4 Hydrogen against Claude Haiku 5.5 for prompts under 100,000 tokens, checked Nov. 14, 2026
Each row compares one kind of token. The blue dot is Claude Haiku 5.5's price, the magenta dot is Gemini 4 Hydrogen's price, and the rod shows the gap between them.
Sources: Google [1]; Anthropic [2]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Claude Haiku 5.5 | Gemini 4 Hydrogen |
|---|---|---|
| Input tokens | $0.1 | $1.5 |
| Output tokens | $0.5 | $7.5 |
Artificial Analysis benchmarks AI models on its own3. It tested Hydrogen the day it shipped. One task costs $0.29 at Hydrogen's lowest reasoning setting.
The same task costs $0.80 at medium, Hydrogen's second setting3. Raising it to high, the default, lifts the price to $1.87. The top setting, xhigh, costs $3.16, nearly eleven times the lowest setting's price.
Box Runs It on Real Work
Box tested Hydrogen before the public launch5. Aaron Levie, the company's co-founder and chief executive, posted the results on X. "Ok this is a big deal. We've been testing Gemini 4 Hydrogen in early access with the Box AI Complex Work Eval, and it's a serious jump for fast models," Levie wrote.
Box's own agent scored five percentage points higher with Hydrogen than with Flash, Levie said5. It finished the same work in under half the time. The agent also used about 12 percent fewer tokens. Box's eval runs real documents from finance, consulting and life sciences firms, Levie wrote.
Vasu Deo Sankrityayan reviews AI tools for Analytics Vidhya. He tested Hydrogen by hand and found a similar lift on a smaller task7. His test asked the model to write a SQL query against a sample set of game rules.
"At Medium thinking, it handled rule lookup and produced SQL that held up beyond the supplied records," he wrote. Box plans to add Hydrogen to Box AI Studio soon, Levie said. Any customer will then be able to build their own agent on it5.
The Case Against Switching Today
Artificial Analysis found a weaker side too3. "Hydrogen scores 54% on Terminal-Bench 4.0 and 50% on Terminal-Bench-Science 0.1, covering agentic coding and scientific research workflows in the terminal," the firm wrote. Claude Sonnet 5.5 scored higher on a related test. At its top setting, it reached 63.9 percent on Terminal-Bench 4.1 that same week, VentureBeat reported4.
Artificial Analysis@ArtificialAnlysGoogle's new Gemini 4 Hydrogen has Intelligence just below Claude Sonnet 5.5 and GPT-6.1 Sol, at a third of Sonnet’s cost but ~2.5x that of GPT-6.1 Sol@GoogleDeepMind has released their third new model in six weeks. Gemini 4 Hydrogen introduces an xhigh reasoning level and… pic.twitter.com/mtn4GzkVVq
A reader doing heavy terminal work pays more for Hydrogen. The model still trails a pricier Anthropic rival on that specific test. Running simple, high-volume jobs instead, a buyer pays fifteen times what Haiku charges. Haiku may already handle that work well enough.
Plenty of developers questioned the whole exercise of switching again. Hannan Shoukat founded the studio 8Bit Enterprise. He wrote about the choice days before Hydrogen shipped.
"My first reaction wasn't excitement. It was: do I really switch again?" Shoukat wrote. Benchmarks climb most weeks, he added, while his daily client work stays steady8.
Hydrogen's Task Cost Spans $0.29 to $3.16 by Reasoning Level
Artificial Analysis's estimated cost per Intelligence Index task, Gemini 4 Hydrogen by reasoning level, Nov. 12, 2026; lower is better
Each column is one reasoning setting, low to xhigh. The height is the dollar cost Artificial Analysis measured to finish one benchmark task at that setting, so a shorter column costs less.
Source: Artificial Analysis on X [3]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Low | $0.29 |
| Medium | $0.80 |
| High | $1.87 |
| Xhigh | $3.16 |
Hydrogen also drops a feature some developers already built around. Google Cloud's migration guide tells teams to delete their explicit caching calls6. It tells them to lean on Hydrogen's automatic caching instead. A team moving code from Flash has to rewrite that part of its setup first.
What to do
Match the model to the job before you switch anything. Running single, high-stakes tasks where a mistake costs real money, check your own cost per finished task too3. Compare that figure alongside the price per token, the way Artificial Analysis does. Hydrogen's extra accuracy and speed may earn its premium there.
Running high-volume, routine jobs, test Claude Haiku 5.5 first. It costs a fraction of Hydrogen's price for short prompts2. Box's own accuracy gain over Flash ran just five points5.
Hold off migrating existing Flash code. Rewrite any explicit caching calls first. Google Cloud's guide walks through the change end to end6.
By the numbers
- $1.50 and $7.50: Hydrogen's price per million input and output tokens, flat across its full 1 million-token window1.
- 15 times: how much more Hydrogen charges per token than Claude Haiku 5.5 for a prompt under 100,000 tokens2.
- 57 percent: Hydrogen's score on Terminal-Bench 4.1, against 18.2 percent for Gemini 3.8 Flash1.
- Five percentage points: Box's accuracy gain switching its own agent from Flash to Hydrogen5.
- $1.87: the cost Artificial Analysis measured for one benchmark task at Hydrogen's default setting, against $3.16 at its highest setting3.
- 63.9 percent: Claude Sonnet 5.5's score on the related Terminal-Bench test, ahead of Hydrogen's 57 percent4.
- Explicit caching: the feature Google Cloud's migration guide tells developers to replace with automatic caching when they move off Flash6.
- Nov. 12, 2026: the launch date for Hydrogen across the Gemini API, AI Studio and the free Gemini app's automatic router1.
What to watch
Watch whether Google adds a cheaper Hydrogen tier. Flash eventually got one. Box's five-point accuracy gain still needs a check once Hydrogen leaves early access.
Artificial Analysis tends to publish a second benchmark run within weeks of a launch. A follow-up test usually settles whether a launch-day number holds up.
Sources
- Koray Kavukcuoglu, "Introducing Gemini 4 Hydrogen," Google, Nov. 12, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/
- Anthropic, "Pricing," Anthropic, accessed Nov. 14, 2026, https://platform.claude.com/docs/en/about-claude/pricing
- Artificial Analysis, "Google's new Gemini 4 Hydrogen has Intelligence just below Claude Sonnet 5.5 and GPT-6.1 Sol," X, Nov. 12, 2026, https://x.com/ArtificialAnlys/status/2120941511223185722
- Carl Franzen, "Google launches Gemini 4 Hydrogen — a faster, smarter, pricier Flash successor," VentureBeat, Nov. 12, 2026, https://venturebeat.com/technology/google-launches-gemini-4-hydrogen-a-faster-smarter-pricier-flash-successor
- Aaron Levie, "Ok this is a big deal," X, Nov. 12, 2026, https://x.com/levie/status/2120947988143198302
- Google Cloud, "Developer's guide to Gemini 4 Hydrogen," Google Cloud, Nov. 12, 2026, https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-hydrogen
- Vasu Deo Sankrityayan, "Gemini 4 Hydrogen Review: Performance and Practical Tests," Analytics Vidhya, Nov. 13, 2026, https://www.analyticsvidhya.com/blog/2026/11/gemini-4-hydrogen-review/
- Hannan Shoukat, "Gemini 4 Hydrogen Dropped This Week. Which AI Model Do You Actually Code With Every Day, and Why?" DEV Community, Nov. 12, 2026, https://dev.to/hannanshoukat/gemini-4-hydrogen-dropped-this-week-which-ai-model-do-you-actually-code-with-every-day-and-why-fcn
