On November 12, 2026 (US time), Google released Gemini 4 Hydrogen. It is the model in Gemini 4 whose job is to get work done quickly, and Google's announcement calls it a "daily driver" for agentic work. The job of thinking long and hard on difficult problems belongs to Gemini 4 Carbon and Argon, which came out earlier.
Key points
- The model
- A speed-focused successor to Gemini 3.8 Flash. According to Google's announcement, at its default settings it finishes tasks 2.5 times faster than 3.8 Flash.
- Best for
- Everyday coding, agents that work by calling tools, and high-volume processing.
- Price
- The API has a free tier. On the paid tier, it costs $1.50 input and $7.50 output per million tokens. That is cheaper than Carbon and Argon ($4 and $20 from December 1), and more expensive than 3.8 Flash's limited-time price through December 31 ($0.75 and $3.75, doubling on January 1, 2027).
Checked November 13, 2026, against Google's announcement, the Gemini API documentation, and the pricing page
Hydrogen = fast execution
- Everyday coding, fixes, and running tests
- Agents that call tools, and parallel subagents
- $1.50 input and $7.50 output, with a free tier
Hydrogen implements and verifies
Carbon and Argon = long reasoning
- Design reviews, bugs with unknown causes, and research
- Longer waits, because each response involves deep thinking
- $4 input and $20 output from December 1, no free tier
Based on Google's announcement (November 12, 2026) and the Gemini Developer API pricing page (checked November 13, 2026). The differences between Carbon and Argon are covered in detail in Gemini 4 Carbon vs. Argon.
1. What is Gemini 4 Hydrogen? The "fast" model in Gemini 4
As Google describes it, Hydrogen is a model built to work quickly and reliably on everyday coding and agentic tasks. The model card lists its intended uses as fast execution, either on its own or as a subagent working under a more capable model; iterative coding and debugging in a repository; open-ended agentic tasks; knowledge work across documents and images; and writing, research, and prototyping.
Like Argon and Carbon, it is named after a chemical element, a separate naming line from the Flash and Pro names Gemini has used so far. How the model names in the Gemini app have changed, and how this model relates to other Gemini models, are laid out in What is Google Gemini?
| Item | Gemini 4 Hydrogen |
|---|---|
| Model ID | gemini-4-hydrogen (a single stable version) |
| Release | Generally available (GA) on November 12, 2026 |
| Context window | 1,048,576 tokens for input, thinking, and output combined |
| Output per request | Up to 65,536 tokens normally; with background execution, close to 1 million tokens minus the input |
| Thinking levels | low, medium, high (default), xhigh |
| Input types | Text, images, audio, video, and PDF (output is text only) |
| Supported tools | Function calling, structured output, grounding with Google Search and Google Maps, code execution, computer use, file search, and URL context |
| Not supported | Audio generation, image generation, the Live API, and explicit caching |
| Knowledge cutoff | March 2026 (earlier in some areas) |
| API free tier | Yes (Carbon and Argon have none) |
Sources: Gemini 4 Hydrogen model page; Gemini 4 Hydrogen model card; What's new in Gemini 4 Hydrogen (checked November 13, 2026). Cutoffs for models from each company are collected in AI knowledge cutoff dates.
2. When to use Hydrogen, Carbon, Argon, or 3.8 Flash
Gemini 4 has three models, and the previous generation's 3.8 Flash is still offered. Lining up Google's positioning and prices shows two axes for choosing: how deeply you want the model to think on each response, and speed and cost.
| Model | Google's positioning | Input / output | Default / max thinking | API free tier |
|---|---|---|---|---|
| Gemini 4 Hydrogen | The workhorse for everyday coding and agentic tasks, and parallel subagent execution | $1.50 / $7.50 | high / xhigh | Yes |
| Gemini 4 Carbon | The most capable model, for design, debugging, science, and security audits | $2 / $10$4 / $20 from December 1 | medium / high | No |
| Gemini 4 Argon | Most advanced long-horizon tasks, with fewer tokens than Carbon | $2 / $10$4 / $20 from December 1 | medium / high | No |
| Gemini 3.8 Flash | The previous generation's speed-focused model (Hydrogen is the migration target) | $0.75 / $3.75$1.50 / $7.50 from January 1, 2027 | medium / high | Yes |
Per million tokens, in US dollars, at Standard processing. Output includes thinking tokens. Sources: Gemini Developer API pricing; the model selection guide in What's new in Gemini 4 Hydrogen (checked November 13, 2026). 3.8 Flash has three thinking levels, low, medium, and high (that medium is its default comes from Hydrogen's migration guide).
The price relationship is worth a close look. 3.8 Flash's price is a limited-time price through December 31, 2026, and when it doubles on January 1, 2027, its input and output prices become the same as Hydrogen's (cache reads for 3.8 Flash become $0.15 from January, versus $0.075 for Hydrogen). Compared with Carbon and Argon, from December 1 Hydrogen's input and output prices are both three-eighths of theirs (just under 40%).
- HydrogenFirst choice
Everyday coding, adding features, and running tests Work where an agent calls tools and gets things done. It responds quickly, and its price is the lowest in Gemini 4.
- HydrogenParallel subagents
Implementer and checker working under a more capable model Google recommends a setup where Carbon acts as the coordinator and Hydrogen handles the work and the verification.
- CarbonOnly the hard parts
Design reviews, bugs with unknown causes, and reading long documents Escalate tasks to it once you find that even Hydrogen's
xhighisn't enough. - ArgonStandard long reasoning
Advanced long-horizon tasks, with fewer tokens than Carbon Same price as Carbon, with less thinking per response.
- 3.8 FlashUntil December 31
High-volume processing where price comes first While its limited-time price lasts, it costs half as much per token as Hydrogen. If speed and cost matter above all, there is also 3.5 Flash-Lite, which Google describes as its fastest and cheapest model.
Combining Carbon and Hydrogen
Google Antigravity has an orchestrator that switches between Carbon and Hydrogen within a single task. According to Google's announcement, on Terminal-Bench 4.0 it came close to the success rate of Carbon alone while using 24% less quota and running 19% faster on average (Google's own measurements). For parallel work, Google also shows a setup where Carbon acts as the coordinator and Hydrogen runs worker and verifier agents in separate workspaces. To build the same setup with the API, assign planning to Carbon, and file edits, subagent tasks, and test runs to Hydrogen.
3. Pricing, and a sample cost calculation for the same workload
| Tier | Input | Output (including thinking) | Cache reads |
|---|---|---|---|
| Free tier | Free | Free | Free |
| Standard | $1.50 | $7.50 | $0.075 |
| Batch and FlexNon-urgent processing | $0.75 | $3.75 | $0.0375 |
| PriorityProcessed with priority | $2.70 | $13.50 | $0.135 |
| ContributorInvitation-only preview; content is used for product improvement | $0.75 | $3.75 | $0.0375 |
Per million tokens, in US dollars. The same price applies to long contexts. Source: Gemini Developer API pricing (checked November 13, 2026). Grounding with Google Search and with Google Maps is each free for up to 5,000 queries a month (shared across all Gemini 3 and later models), then $14 per 1,000 queries. Grounding isn't available on the free tier.
Caching is implicit caching only. It applies automatically to repeated prefixes of 4,096 tokens or more, and input that hits the cache is 95% off. Creating caches yourself (explicit caching) is not supported. When you use grounding such as Search, the tokens for instructions and function definitions that Google adds behind the scenes are also billed.
Sample calculation: 1,000 requests a day for a month
Assumptions (set for this article)
1,000 calls a day, each with 20,000 input tokens and 2,000 output tokens (including thinking), for 30 days. That is 20 million input tokens and 2 million output tokens a day. We assume every model uses the same number of tokens and compare prices only.
| Model and tier | Per day | 30 days |
|---|---|---|
| Hydrogen, Standard | $45.00 | $1,350 |
| Hydrogen, Batch/Flex | $22.50 | $675 |
| Hydrogen, Standard + implicit cachingWhen 15,000 input tokens hit the cache every time | $23.63 | $708.75 |
| 3.8 Flash, StandardUntil December 31, 2026 | $22.50 | $675 |
| 3.8 Flash, StandardFrom January 1, 2027 | $45.00 | $1,350 |
| Carbon/Argon, StandardUntil November 30, 2026 | $60.00 | $1,800 |
| Carbon/Argon, StandardFrom December 1, 2026 | $120.00 | $3,600 |
Breakdown for Hydrogen, Standard: input 20 million × $1.50 ÷ 1 million = $30, and output 2 million × $7.50 ÷ 1 million = $15. The caching row prices the 15 million cached tokens at $0.075 and the remaining 5 million at $1.50. If every call is grounded with Google Search, $350 is added for the 25,000 queries left after the 5,000 free ones out of 30,000 a month.
The table shows two things. First, sending non-urgent work to Batch or Flex, or simply putting the same prefix (instructions and documents) at the start of every request so it hits implicit caching, cuts the cost roughly in half. The cache only hits when the prefix is identical every time, so put the parts that change (such as the user's question) at the end. Second, the gap with Carbon and Argon widens from December 1: for the same number of tokens, Hydrogen costs three-eighths as much (just under 40%).
Your actual cost is the price multiplied by the number of tokens that model uses. The table above assumes equal token counts; in practice, the amount of output (including thinking) varies by thinking level and by model. Hydrogen defaults to high, and xhigh uses even more. In a third-party measurement on the same evaluation tasks, Hydrogen (xhigh) produced 304 million tokens and 3.8 Flash (high) produced 126 million (section 5).
Run 10 to 20 of your own tasks on each model, work out the cost per task from the input, output, thinking, and cache token counts in the response's usage_metadata, and then decide.
The full picture of Gemini's free version, subscriptions, and API pricing is in Gemini pricing and what you can do for free.
4. The API free tier, and how your data is handled
Hydrogen is available on the Gemini API free tier. It is the only Gemini 4 model with a free tier; Carbon and Argon are paid tier only. You can also try it for free in Google AI Studio.
Content you send on the free tier is used to improve Google's products. On the pricing page, the row for whether content is used to improve products says "Yes" for the free tier and "No" for the paid tier. The invitation-only Contributor preview tier is half the Standard price, but its content is also used for product improvement.
If you handle confidential documents, customers' personal information, or unreleased code, call the API on the paid tier (with an API key that has billing set up). It is safest to limit the free tier to prototyping and testing with content that can be made public.
The Google Cloud credits that come with Google AI Pro and Ultra each month can be applied to paid-tier charges. How to claim them is covered in API credits included with subscriptions.
5. Performance: Google's published numbers and third-party measurements
Google's published numbers
The numbers below were all published by Google itself. They come from evaluations and conditions chosen by Google, which differ from other companies' announcements.
| Evaluation | Hydrogen | Compared with |
|---|---|---|
Task completion speedHydrogen at the default high | 2.5 times 3.8 Flash | 3.8 Flash |
| Terminal-Bench 4.1Terminal tasks | 57% | 3.8 Flash 18.2% |
| Artificial Analysis Intelligence IndexComposite score | high 51 / xhigh 54 | Argon 50 |
| HLE-DiamondReasoning on hard problems | 58.8% | Top in its price range (per Google) |
| Vals Tax Agent BenchTax | 52.3% | First overall (per Google) |
| Vals Legal Research BenchLegal research | 55.8% | Tied for first (per Google) |
| MKAKnowledge work across images and documents | 53.4% | Ahead of other models (per Google) |
| Agents' Last Exam v1.1Agentic knowledge work | 71.6% | — |
Source: Gemini 4 Hydrogen announcement (Google, November 12, 2026). The 2.5 times figure refers to the time to finish a task compared with 3.8 Flash (the settings used for 3.8 Flash aren't stated). Google says early-access companies saw 2 to 2.8 times faster performance and 12 to 23% fewer tokens on chart, SQL, and data-heavy tasks.
The model card also has a table comparing Hydrogen with other companies' models in a similar price range (Claude Sonnet 5.5 and GPT-6.1 Terra). There are evaluations where Hydrogen comes out ahead and others where it falls behind.
| Evaluation | Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweepAgentic coding | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1Adjusted for cheating | 57.0% | 63.9% | — |
| Agents' Last Exam v1.1 | 71.6% | 60.4% | 69.8% |
| AutomationBench-AAWork automation | 76.1% | 71.8% | 60.2% |
| OSWorld-2.1Computer use (Offline subset) | 77.6% | 83.9% | 74.2% |
| MKA | 53.4% | 52.0% | 43.8% |
| ChartographyReading charts | 66.6% | 61.6% | 63.2% |
| Terminal-Bench-Science 0.1Scientific tasks | 51.4% | 59.9% | 58.3% |
| HLE-Diamond | 58.8% | 53.3% | 47.3% |
Source: Gemini 4 Hydrogen model card (November 2026 values; checked November 13, 2026). The methodology is on the evaluation methodology page linked from the model card. Only some rows of the table are shown.
Claude Sonnet 5.5 came out ahead on terminal tasks, computer use, and scientific tasks, while Hydrogen led on knowledge work, work automation, and reasoning on hard problems. Hydrogen is not the best at every task. If you are choosing a coding tool, the reliable approach is to try both on tasks in your own repository.
On safety, Google says that compared with 3.8 Flash, Hydrogen oversteps into actions it wasn't asked to take up to 81% less often, and its tendency to try to cross permission boundaries when it can't complete a task is down by as much as half (Google's evaluations).
Third-party measurements (Artificial Analysis)
These values are from Artificial Analysis, which measures models independently. It publishes a composite score that combines several evaluations (Intelligence Index v5.0.1), along with the cost and time per task on those evaluations. Note that the settings measured differ by model (xhigh for Hydrogen, high for the others).
Source: Artificial Analysis model pages (Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Gemini 3.8 Flash; checked November 14, 2026). Costs are calculated at each model's current API price (Carbon and Argon at their price through November 30, 3.8 Flash at its price through December 31). Values change when models are re-measured.
Three things stand out. First, its composite score is above Argon's and 6 points short of Carbon's. Second, its cost per task was the lowest of the four, and its time per task was the shortest. 3.8 Flash has the lower price per token, and Hydrogen also produced more output tokens (304 million versus 126 million), yet Hydrogen's total per task was less than half. Both used more than 10 billion input tokens (11.3 billion for Hydrogen and 10.2 billion for 3.8 Flash), and the cache discount differs (95% for Hydrogen, 90% for 3.8 Flash), so most of the cost appears to be driven by input and how often it hits the cache. The cost breakdown hasn't been published. Third, Google's claim that Hydrogen is 2.5 times faster than 3.8 Flash is a comparison with Hydrogen at its default high, and the settings for 3.8 Flash aren't stated. Between Artificial Analysis's xhigh Hydrogen and high 3.8 Flash, time per task was almost the same (14.3 and 14.6 minutes), and 3.8 Flash was faster to the first token (16.5 versus 44.9 seconds). If you want the first words of a chat reply to arrive quickly, start by lowering the thinking level and measuring.
6. Choosing a thinking level, from low to xhigh
Hydrogen has four thinking levels, selected with thinking_level in generation_config. The default is high, and within Gemini 4 only Hydrogen has xhigh (Carbon and Argon default to medium, with a maximum of high).
low
Where cost and speed matter most, such as quick chat and high-volume processing (fewest tokens and shortest waits)
medium
Interactive coding, UI prototyping, and tool calls where latency matters
highDefault
The recommended level for most tasks: complex coding, agent runs, tool calls, and work across images and documents
xhigh
Only for difficult, many-step coding and agentic tasks where high falls short (it overthinks simple questions)
Source: What's new in Gemini 4 Hydrogen (checked November 13, 2026). Specifying minimal returns an error.
Google's recommended approach is to set a generous output limit (max_output_tokens), try all four levels on your own evaluation problems, and choose the lowest level that meets your quality bar. In practice, working in the following order means less rework.
- Pick 10 to 20 representative problems from your own work, and decide in advance on the correct answers or pass criteria.
- Run them at the default
high, and record how many pass, along with tokens and time per problem. - Run them at
mediumandlowas well, and find the lowest level where the pass count doesn't drop. - Run only the problems that failed at
highatxhigh. If they still fail, send that type of task to Carbon.
7. Steps for moving from 3.8 Flash to Hydrogen in the API
Google recommends Hydrogen as the migration target from 3.8 Flash and from earlier Gemini 3 and 2.5 models. Changing the model ID alone isn't enough, so work through the items on the migration guide's checklist in order.
- Change the model ID from
gemini-3.8-flashtogemini-4-hydrogen. - Review your output limit. On 3.8 Flash, the 1 million input tokens and 64K output tokens were separate limits, but Hydrogen shares a single window across input, thinking, and output. Remove a low
max_output_tokenssuch as 8,192 or raise it to 65,536, and if you want less output, lower the thinking level. - Decide on
thinking_level. The default changes frommediumtohigh, so comparemediumandhighon your own evaluation problems. Remove anyminimalsetting. - Remove explicit caching (
cachedContents,caches.create, andcached_contentin requests) and rely on implicit caching. - Remove
temperature,top_p, andtop_k(they are ignored if sent). - Declare tools properly with function calling. If you only write "call tools in JSON" in the system instructions, you may get
MALFORMED_FUNCTION_CALL. - Preserve conversation state. With the Interactions API, use
previous_interaction_id; if you manage history yourself with GenerateContent, send back the previous model responses and their thought signatures unchanged. - If you need more than 64K tokens of output, set
backgroundto true and setmax_output_tokensabove 65,536. - Redo your cost estimates. Thinking tokens are billed, as are the instructions and function definitions Google adds when you use grounding such as Search.
Source: Migration checklist in What's new in Gemini 4 Hydrogen (checked November 13, 2026). Problems caused by the automatic redirection of Gemini 3 model names are covered in Gemini Flash and Pro auto-redirects and migration.
Differences when switching from Carbon or Argon to Hydrogen
| Item | Carbon and Argon | Hydrogen |
|---|---|---|
| Request format | The same (no changes needed) | |
| Default thinking | medium | high |
| Max thinking | high | xhigh |
| Output per non-streaming call | 32,768 tokens (REST) | 65,536 tokens |
| API free tier | No | Yes |
Source: What's new in Gemini 4 Hydrogen (checked November 13, 2026). You can also move only the work where speed comes first to Hydrogen, and keep the work that needs deep thinking on Carbon or Argon.
In Gemini Managed Agents, the default model for Antigravity agents is now Hydrogen. To use a different model, change it with agent_config.
8. Where you can use it: the API, the app, Search, and Sheets
Where you can use Gemini 4 Hydrogen (checked November 14, 2026)
Gemini API and Google AI StudioBoth the free tier and the paid tier (GenerateContent and the Interactions API)
AvailableGemini appSelectable in the model picker, and used for free users through automatic model selection when it suits the question
AvailableGoogle SearchAvailable to all users worldwide (per Google)
AvailableGoogle AntigravityPaid Google AI Pro and Ultra subscribers
Pro and UltraGemini in Google SheetsRolling out starting with Google AI Pro and Ultra subscribers, plus eligible business plans
Pro and UltraGemini for business and Gemini EnterpriseIn the API, the model picker, Antigravity, and Sheets (eligible plans)
Business plansAndroid StudioAvailable to developers
AvailableGitHub CopilotCopilot Pro, Pro+, Max, Business, and Enterprise, from the model picker in VS Code, Copilot CLI, the coding agent, and more (rolling out)
Paid plansCursorAdded to the list of supported models (at Cursor's pricing)
Available
Sources: Gemini 4 Hydrogen announcement; Google Search blog post (both November 12, 2026); Gemini API release notes (checked November 13, 2026); GitHub Changelog (November 13, 2026); Cursor documentation (checked November 14, 2026).
In GitHub Copilot, Hydrogen is now generally available, while Gemini 3.8 Flash will be removed from every Copilot experience on December 14, 2026 (GitHub's announcement; Hydrogen is the replacement). If you use 3.8 Flash in Copilot, switch before then.
For Google Search, Robby Stein, who leads Search at Google, describes this as Search's biggest model upgrade to date. Google says it improves understanding of multi-part questions and typos, the process of researching related subtopics at the same time to surface a wider range of links, and answers that use images and comparison tables, and that answers improve in every language AI Mode supports.
In the Gemini app, Carbon is limited to Google AI Ultra subscribers (Gemini 4 Carbon vs. Argon), whereas Hydrogen is used even for free users through automatic selection when it suits the question. Antigravity and Sheets are available from the paid Pro plan. Google says output in the Sheets interface is more than 30% faster than with previous models. For each plan's limits, see Gemini pricing and what you can do for free.
Summary
- Gemini 4 Hydrogen, generally available since November 12, 2026, is the "fast execution" model in Gemini 4. Hard problems that need long reasoning go to Carbon and Argon.
- The API price is $1.50 input and $7.50 output, and there is a free tier. That is three-eighths (just under 40%) of Carbon's and Argon's price from December 1.
- Content sent on the free tier is used for product improvement. Use the paid tier for confidential content.
- According to Google's announcement, it is 2.5 times faster than 3.8 Flash and scores 57% on Terminal-Bench 4.1. The model card also shows evaluations where Claude Sonnet 5.5 is ahead.
- In third-party measurements, its composite score is 54 (
xhigh), above Argon and below Carbon. Its cost and time per task were the lowest of the four Gemini models. - The default thinking level is
high. Try all four levels on your own problems and pick the lowest one that meets your bar. - When moving from 3.8 Flash, review the output limit, the default thinking level, explicit caching, and
temperature.
FAQ
Q. What is Gemini 4 Hydrogen?
It is a Gemini 4 model that Google released on November 12, 2026, built to handle everyday coding and agentic tasks quickly. It is positioned as the successor to Gemini 3.8 Flash, while hard problems that need long thinking go to Carbon and Argon, the other Gemini 4 models.
Q. How much does the Gemini 4 Hydrogen API cost?
It costs $1.50 input and $7.50 output per million tokens (output includes thinking tokens). Batch and Flex are half that, and input that hits implicit caching is $0.075. It is also available on the Gemini API free tier.
Q. Should I use Hydrogen or Carbon and Argon?
For speed and cost, the safe approach is to try Hydrogen first for everyday coding and tool-calling work, and to escalate to Carbon only for design reviews or bugs with unknown causes where even xhigh isn't enough. Google also recommends a setup with Carbon as the coordinator and Hydrogen as the worker and verifier.
Q. Can I use Gemini 4 Hydrogen for free?
Yes. You can try it on the Gemini API free tier and in Google AI Studio, and in the Gemini app it is used even for free users through automatic model selection when it suits the question. However, content you send on the API free tier is used to improve Google's products.
Q. What changes when I switch from 3.8 Flash?
The default thinking level changes from medium to high, and the output limit becomes part of a single window shared with input. Explicit caching and parameters such as temperature aren't supported, so remove them. The price is double 3.8 Flash's limited-time price through December 31, and from January 2027, when 3.8 Flash's price doubles, the input and output prices are the same.
Update history
- November 13, 2026First published.
Sources
- Gemini 4 Hydrogen (announcement) (Google, Koray Kavukcuoglu, November 12, 2026)
- Gemini 4 Hydrogen and Google Search (Google, Robby Stein, November 12, 2026)
- What's new in Gemini 4 Hydrogen (Gemini API)
- Gemini 4 Hydrogen model page (Gemini API)
- Gemini Developer API pricing
- Gemini API release notes (entry for November 12, 2026)
- Gemini 4 Hydrogen model card (Google DeepMind)
- Gemini 4 Hydrogen (Artificial Analysis)