Gemini 4 Hydrogen pricing, plans and limits
The execution model in Google's Gemini 4 family, meant to run edits, tests and tool calls beside Carbon and Argon.
- Best for
- Engineering teams running coding agents
- Not for
- Hard architecture and security audits
Should you use Gemini 4 Hydrogen?
Google DeepMind released Gemini 4 Hydrogen on 12 November 2026; Artificial Analysis scores it 54 at xhigh, and Google's card reports 76.1% on AutomationBench-AA. It suits teams running everyday coding agents at a lower token price than Carbon or Argon, while the hardest reasoning belongs on Carbon.
Best for
- Engineering teams running coding agents Google built Hydrogen as the fast worker that edits code, runs tests and verifies changes beside Carbon, and its migration notes let both share one session.
- Developers prototyping long-context, multimodal apps without paying first The free tier covers the Gemini API and AI Studio within rate limits, and the shared window takes PDFs, video, audio, images and plain text.
- Knowledge workers reading dense PDFs, charts and recorded video Google reports 53.4% on its Multimodal Knowledge Agent benchmark and 66.6% on Chartography, both vendor-reported and taken from a table Google chose.
Not for
- Hard architecture and security audits Google says Carbon spends more thinking tokens per turn to test hypotheses and verify edge cases on hard architecture, debugging, science and security audits.
- Voice agents and live conversation products The Live API is not supported, and Artificial Analysis measured a 16.68 s wait for the first answer token at the default high level.
- Workloads that need image or audio output Hydrogen returns text only.
Can you use Gemini 4 Hydrogen?
- Gemini API and Google AI Studio: yes (Free tier (rate-limited) and paid tier; Batch and Flex are paid only; 230 countries). Free-tier content may be used to improve Google products, except in the EEA, Switzerland and the UK.
- Gemini app: yes (Model drop down; free users get it through Auto model selection when it fits the prompt).
- Google Search: yes (All users).
- Google Antigravity: yes (Google AI Pro and Ultra subscribers).
- Gemini in Google Sheets: yes (Google AI Pro and Ultra subscribers, and eligible business subscriptions).
- Free plan: yes.
Checked 13 Nov against Google DeepMind's own pages.
Gemini 4 Hydrogen pricing
- Input, per 1M tokens
- $1.50/1M
- Output, per 1M tokens
- $7.50/1M
- Cached input, per 1M tokens
- $0.07/1M
- Pricing model
- per-token
- Free tier
- Free of charge within Google's rate limits; the Batch and Flex modes are paid-tier only.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.045 | $0.0075 | $0.052 |
| Support reply | $0.0030 | $0.0022 | $0.0052 |
| One coding agent run | $0.300 | $0.150 | $0.450 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Billed per token on the paid tier, and the output price covers thinking tokens as well as the visible answer. Google counts every token the model processes, including internal system instructions and function declarations when grounding is on. The rate is flat across the full window, so a long prompt costs the same per token as a short one. Cached input is implicit only and gets a 95% discount on a hit. Grounding with Google Search or Maps includes 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000 requests, with Search capped at 1,500 requests a day.
What Gemini 4 Hydrogen can do
Gemini 4 Hydrogen is Google DeepMind's fast multimodal model, scoring 51 on the Artificial Analysis Intelligence Index at its default high thinking level. It accepts PDFs, video, audio, images and plain text inside one shared 1,048,576-token window and writes text only.
What goes in and what comes out
Reads
- text
- image
- audio
- video
Writes
- text
Capabilities
- Audio inputSupported
- Code executionSupported
- Computer useSupported
- File searchSupported
- Function callingSupported
- Long decode continuationSupported
- Maps groundingSupported
- Search groundingSupported
- Structured outputSupported
- Tool useSupported
- URL contextSupported
- Video inputSupported
- VisionSupported
- Audio outputNot supported
- Image generationNot supported
- Live APINot supported
- Context caching
- implicit only
- Reasoning modes
- low, medium, high, xhigh
Key features
Four thinking levels
Set thinking_level to low, medium, high (the default) or the new xhigh; minimal is rejected with an error.
Long decode past 64K
Background execution lets one request run beyond a single 64K hop, up to the shared window minus the input.
Parallel tool batching
Batches file reads and searches in parallel instead of one by one; Google reports partners saw runs 2x to 2.8x faster with 12% to 23% fewer tokens than Gemini 3.8 Flash.
Multimodal inputs, text output
Takes text, images, video, audio and PDFs and answers in text, with agentic video understanding that Google says typically needs up to 88% fewer input tokens on long video.
Show all 7 features
Built-in tools
Function calling, structured outputs, code execution, computer use, file search, URL context and grounding with Google Search and Maps are all listed as supported.
Implicit context caching
Runs automatically on prompts of at least 4,096 tokens and discounts cached input tokens by 95% on a hit; explicit caching is not supported.
Default model for Antigravity's managed agent
The Antigravity agent in Gemini Managed Agents now uses Gemini 4 Hydrogen by default.
Known limitations
- Explicit context caching is not supported; only automatic implicit caching applies.
- The temperature, top_p and top_k settings are ignored by the server.
- The knowledge cutoff is March 2026, and in some domains the model's knowledge is less current.
- Output past 64K tokens needs background execution or client-side continuation.
Who it is for
- Software engineers running coding agents and parallel sub-agents
- Developers prototyping long-context, multimodal apps on a free tier
- Teams analysing documents, charts and video in a single request
- Enterprise buyers using Gemini for business
Gemini 4 Hydrogen benchmarks
Gemini 4 Hydrogen against its nearest rivals
Output speed tokens per second, higher is better
Standout result
Artificial Analysis Intelligence Index v5.0.1 at the default high thinking level
Source: Artificial Analysis model page, read 13 November 2026Where it sits
Priced around the middle of the 104 GA models with a published price (rank 61), and rank 16 of 58 on output speed as cited from Artificial Analysis.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $3.00/M | #61 / 104 | peer median $1.73/M |
| Output speedtokens/s | 179 tok/s | #16 / 58 | peer median 95 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Measured figures
- LMArena Elo
- 1490
- LMArena rank
- #15
- Output speed
- 179 tok/s
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- Free to try: Google lists it on the free tier of both AI Studio and the Gemini API, and the Gemini app serves it to free users through Auto selection when it fits the prompt.
- Fast on an independent test: Artificial Analysis measured 179 output tokens per second at the default high level.
- Google reports that fabricated claims fell by more than five times against Gemini 3.8 Flash because it flags missing context instead of guessing; this is a vendor-reported figure.
- Carbon and Hydrogen turns can share one session by passing previous_interaction_id, and Google's migration notes say moving between them needs no request schema changes.
Show 2 more
- Per Google's model page, read on 13 November 2026, input, thinking and output share one window, so long video, large PDFs and charts can go in a single request.
- Google's thinking-levels guide, read on 13 November 2026, describes low for fast chat and high-throughput work and xhigh for hard multi-step agent problems, all on the same model.
Limits
- Very verbose at xhigh: Artificial Analysis counted 304M output tokens to run its Intelligence Index at xhigh against 110M at high, with a 40.55 s wait for the first answer token against 16.68 s, read on 13 November 2026, so xhigh costs more and runs slower.
- Text output only: there is no audio generation, no image generation and no Live API, so voice agents and image tools need another model.
- Google's own automated safety evaluation shows unjustified refusals up 5.57 percentage points against Gemini 3.8 Flash, with the rise most visible on multimodal prompts.
- On the Arena text board its preliminary score is 1490 after 4,741 votes, seven points below Gemini 3.8 Flash (High) at 1497 on the same board, read on 13 November 2026; the gap is smaller than Hydrogen's own margin of nine points.
Gemini 4 Hydrogen alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Gemini 4 Hydrogen This page | 1M | $1.50 / $7.50 | text, image, audio, video, pdf → text | |
 Claude Sonnet 5.5Pick Hydrogen if Google's own table matters for your work, since it leads Sonnet 5.5 on most rows there. Pick Sonnet 5.5 for terminal and desktop-control agents, where that same table has it ahead. | Anthropic | 1M | $2 / $10 | text, image, pdf → text |
 GPT-6.1 TerraPick Hydrogen for coding-agent and knowledge-work scores, where Google's table has it ahead of Terra on every row both report except one. Pick Terra if your work looks like the one science-and-math row where Terra leads. | OpenAI | 1.1M | $1 / $5 | text, image → text |
| Gemini 3.8 FlashPick Hydrogen for new work, which Google names the recommended upgrade. Pick 3.8 Flash only for the months when its token rate is lower, and plan the migration before it changes. | 1M | $0.75 / $3.75 | text, image, video, audio, pdf → text |
Also compared
- Gemini 4 ArgonPick Hydrogen for everyday execution at a lower token price. Pick Argon for the advanced long-horizon tasks Google aims it at.
Gemini 4 Hydrogen specs
- Context window
- 1.0M tokens
- Max output
- 66K tokens
- Tokens per second
- 179 tok/s
- Architecture
- Not disclosed
- Parameters
- Not disclosed
Training data
Google's model card says Gemini 4 Hydrogen is adapted from Gemini 4 Carbon and refers readers to the Carbon card for the training dataset and data processing; its own card gives no training-data detail.
SDKs
- Python
- JavaScript
Prompting patterns
Show 3 prompting patterns
Start at the high thinking level and move to xhigh only when high falls short
thinking_level: "high"
Google's guide says xhigh adds latency and token use and can overthink simple prompts.
Declare tools natively instead of asking for raw JSON in the system prompt
Pass function declarations or a structured-output schema with the request.
Without declared tools, raw-text JSON calls can fail, as Google's migration notes explain.
Evaluate at all four thinking levels with a high max_output_tokens
Run the same suite at low, medium, high and xhigh.
Google's guide says to pick the lowest level that meets your quality bar, and a low cap can cut answers short.
Known quirks
Show 4 known quirks
The output cap counts thinking tokens
max_output_tokens limits thinking and the visible answer together, so a tight cap can cut a response short.
Workaround: Keep the default or raise it, and lower thinking_level to spend fewer tokens.
Raw-text tool calls can fail
Prompting for custom JSON tool calls with zero declared tools can return a MALFORMED_FUNCTION_CALL error.
Workaround: Declare tools through function calling, or constrain the reply with structured outputs.
Background mode stops at one hop
With background set to true and no max_output_tokens, the server stops after the first 64K hop.
Workaround: Set max_output_tokens above the default, up to the window minus your input tokens.
The minimal thinking level is rejected
Hydrogen defaults to high rather than medium and returns an error for minimal.
Workaround: Use low for the cheapest and fastest setting.
Gemini 4 Hydrogen facts
- Deployment
- API · Gemini app
- Licence
- Proprietary (Google Gemini API terms)
- Safety posture
- Balanced
- Released
- 12 Nov 2026
- Current version
- gemini-4-hydrogen
- Family
- Gemini 4
- Variant
- Multimodal
- Data retention
- Paid Gemini API: Google says it does not use prompts or responses to improve its products and logs them for a limited time to detect policy violations; its abuse-monitoring page says prompts, context and outputs are kept 55 days for that purpose. Free tier and unpaid quota: content is used to improve Google products and human reviewers may read it, except in the European Economic Area, Switzerland and the United Kingdom, where the paid-tier terms apply to free use too. No zero-retention option for this model was read.
Content restrictions
- Google's Prohibited Use Policy applies to the Gemini API, and prompts or outputs flagged by its abuse detection can be reviewed by authorized Google staff.
Safety controls
- Usage monitoring
- Yes
- Post-training alignment
- Yes
- Frontier safety framework
- Yes
- Pre-training data filtering
- Yes
- Product level safety filtering
- Yes
Version history
gemini-4-hydrogen · 12 Nov 2026
General availability on the Gemini API, in the Gemini app and Google Search, and in Google Antigravity for paid AI Pro and Ultra subscribers.
gemini-3.8-flash · 02 Sep 2026
Previous Flash release; Google names Hydrogen the recommended upgrade from it.
What changed
Questions about Gemini 4 Hydrogen
How much does Gemini 4 Hydrogen cost per million tokens?
On Google's paid Gemini API, standard use costs $1.50 per 1M input tokens and $7.50 per 1M output tokens, and the output rate covers thinking tokens. Cached input is $0.075 per 1M. Batch and Flex run at $0.75 and $3.75, and Priority at $2.70 and $13.50, all as listed on Google's pricing page.
Can you use Gemini 4 Hydrogen for free, and is it open source?
Yes to free: Google offers it free of charge within rate limits in Google AI Studio and on the unpaid Gemini API quota, and free Gemini app users reach it through automatic model selection. Batch and Flex are paid only. It is not open source: Artificial Analysis lists it as a proprietary model, and Google distributes it only through the channels its model card lists, none of which is a weights download.
Does Google train on my prompts when I use Gemini 4 Hydrogen?
Not on the paid Gemini API, where Google's terms keep prompts and responses out of product improvement. The free tier works the other way: content can be used to improve Google products and reviewers may read it, though users in the EEA, Switzerland and the UK get the paid terms even on free use. For abuse monitoring Google keeps prompts, context and outputs for 55 days.
Gemini 4 Hydrogen vs Gemini 3.8 Flash: what actually improved?
Google reports Terminal-Bench 4.1 rising from 18.2% to 57.0%, a vendor-reported result. The independent picture is closer: both models carry a preliminary tag on the Arena text board, 1490 for Hydrogen and 1497 for Gemini 3.8 Flash. Moving over also means dropping explicit caching and the temperature, top_p and top_k settings. Gemini 3.8 Flash costs half as much per token until 31 December 2026, and from 1 January 2027 Google lists it at Hydrogen's rate.
Gemini 4 Hydrogen vs Claude Sonnet 5.5: which scores higher?
In Google's own model card, Hydrogen leads Claude Sonnet 5.5 on SWE-sweep (13.5% against 3.1%), HLE-Diamond (58.8% against 53.3%) and four other rows. Sonnet 5.5 leads on Terminal-Bench 4.1 (63.9% against 57.0%), OSWorld-2.1 (83.9% against 77.6%) and Terminal-Bench-Science 0.1 (59.9% against 51.4%). Google chose the rows, so these are vendor-reported figures and not a neutral ranking.
About Gemini 4 Hydrogen
Gemini 4 Hydrogen is a multimodal language model and the third in Google DeepMind's Gemini 4 line, after Carbon and Argon. It reached general availability on 12 November 2026, two days after Gemini 4 Carbon, and sits beside Gemini 4 Argon as the faster member of the family. Google's model card says Hydrogen is adapted from Carbon and defers architecture, training data and hardware to Carbon's card, so its parameter count and design are not disclosed. Google calls it the recommended upgrade from Gemini 3.8 Flash and from earlier Gemini 3 and 2.5 models.
Read the full background
Artificial Analysis, an independent evaluator, scores Hydrogen 51 on its Intelligence Index version 5.0.1 at the default high thinking level and 54 at xhigh, and it measured 179 output tokens per second at high (read on 13 November 2026). Google's model card adds vendor-reported results against Claude Sonnet 5.5 and GPT-6.1 Terra: 13.5% on SWE-sweep, a test of finding and repairing bugs without breaking existing code, 71.6% on Agents' Last Exam v1.1, 76.1% on AutomationBench-AA, 58.8% on HLE-Diamond and a $14,549 bank balance on Vending-Bench 2. Google chose that table's rows and rivals, so read it as the vendor's case rather than a neutral ranking.
Input, thinking and output tokens all share one 1,048,576-token window. A single synchronous response stops at 65,536 tokens; for longer output a request turns on background execution with a higher max_output_tokens, and the server continues in 64K hops, which Google calls long decode. Hydrogen allows 65,536 tokens per non-streaming hop, double the 32,768 that Carbon and Argon allow on non-streaming REST calls.
Google positions Hydrogen as the worker in multi-model setups: Carbon plans and untangles hard problems, while Hydrogen runs code edits, tests and verification in parallel. Orchestrator, a feature of Google Antigravity, switches between the two inside one session. On Terminal-Bench 4.0 Google reports near-Carbon task success with 24% less quota used and 19% faster completion, a figure that belongs to the Orchestrator pairing and not to Hydrogen alone.
Developers call it by the model ID gemini-4-hydrogen through the Interactions API or GenerateContent, using the Google Gen AI SDKs for Python and JavaScript or plain REST. Consumers reach it in the Gemini app through automatic model selection on the free plan, and in Google Search for all users.
Google's model card reports that Hydrogen did not reach the critical capability levels for biological information or cybersecurity risk, but did reach the alert threshold for both, and says Google continues to deploy mitigations. As of 13 November 2026 the card still lists the full Safety and Security Report as coming soon.
The case for something else starts with Gemini 4 Carbon, Google's strongest model and the one Google points at hard architecture, debugging and security audits. On Google's own table Claude Sonnet 5.5 is ahead of Hydrogen on three of the ten rows, Terminal-Bench 4.1, OSWorld-2.1 and Terminal-Bench-Science 0.1, so terminal-heavy and desktop-control agents deserve a test there, and GPT-6.1 Terra also leads on Terminal-Bench-Science 0.1. For sub-second latency and bulk classification, Google itself says to keep Gemini 3.5 Flash-Lite.
Sources and guides
Official sources
- Gemini 4 Hydrogen model page Vendor
- Licence text Licence
- System card Safety
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Google DeepMind's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Gemini 4 Hydrogen
Lists that include Gemini 4 Hydrogen
The provider
More AI Models on Hokai
Qwen4-Max-PreviewAlibaba's November 2026 Qwen 4 flagship preview, open only to Token Plan subscribers, with three reasoning levels from low to xhigh.- Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context.
- Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research.
Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens.
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.