Gemini 4 Hydrogen: Coding, Thinking Levels and API Access

Google released Gemini 4 Hydrogen on November 12, 2026 for everyday coding and tool-using agents. It is available in AI Studio and through Google's free and paid Gemini API tiers, on GenerateContent and the Interactions API. The exact model ID is gemini-4-hydrogen, and thinking defaults to high. Start with the access checks below, then test the level on your own tasks.

Where Gemini 4 Hydrogen is available

Google API access details below were checked November 12, 2026 America/New_York (November 13 UTC). Google's launch post and what's-new guide put Hydrogen on the free and paid tiers of the Gemini API and in AI Studio. The model page labels it Stable with a November 2026 update, and the changelog dates general availability to November 12, 2026. API developers can use GenerateContent or the Interactions API.

The free tier is rate-limited and excludes Google Search and Maps grounding, according to Google's pricing page. Google's launch post also names the Gemini app, Search, Gemini for business and Sheets, plus Antigravity for paid Google AI Pro and Ultra subscribers.

OmniaKey now lists Hydrogen as Gemini 4 Hydrogen under the ID gemini-4-hydrogen; its model page carries current OmniaKey capabilities and pricing. The REST example below uses Google's host and documents Google's Gemini API. The thinking-level, long-decode, price and benchmark details are also Google-specific. Interactions, long decode and other Google endpoints are separate capabilities, so check the model page before relying on them. Google's free tier belongs to Google's Gemini API; OmniaKey bills usage separately. For longer multi-step work, Gemini 4 Argon remains listed; Gemini 3.8 Flash covers workflows built around the Flash line. Check the current model list before you switch.

What Hydrogen is built for: coding and tool-using agents

Google positions Hydrogen as a fast, efficient model for everyday coding, interactive web apps, tool-using agents, multimodal document work and parallel subagent tasks. It is the default model for the Antigravity agent in Gemini Managed Agents, and Google recommends it as an upgrade from Gemini 3.8 Flash and earlier Gemini 3 models. Carbon and Argon remain the models Google describes for harder reasoning and longer-horizon work.

Google's launch post also makes vendor claims: cleaner code than Gemini 3.8 Flash, Terminal-Bench 4.1 at 57.0% versus Gemini 3.8 Flash's 18.2%, and tasks finishing 2.5x faster at high. The technical details follow Google's documentation. The performance numbers are Google's report, not results from a new hands-on Hydrogen benchmark. Use them to decide what to test.

Thinking levels: high by default, xhigh for hard tasks

Start at high and reserve xhigh for hard, multi-step coding or agent problems where high falls short. Google's thinking guide lists four levels:

Google does not accept minimal; low is the lightest level. medium is the level to try first when you want faster interactive coding or UI work; low suits high-volume chat and simple processing. xhigh uses the most tokens and can overthink simple extraction, so it is a poor default for routine calls.

The same level name does not mean the same token budget or controlled equal compute across models. Google describes each level by relative token use and does not attach a fixed token count to the label. Hydrogen defaults to high; Carbon and Argon default to medium and stop at high. For an initial comparison, use the same task, output cap and explicit level, but treat that as a starting point: measure total tokens and accepted results, then tune the level and cap for each model.

A REST Interactions example from Google's docs

The following request is copied from Google's thinking guide. It sets low and includes the Api-Revision header shown on that page.

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Api-Revision: 2026-05-20" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemini-4-hydrogen",
    "input": "Provide a list of 3 famous physicists and their key contributions",
    "generation_config": {
      "thinking_level": "low"
    }
  }'

Keep the key in the GEMINI_API_KEY environment variable; the x-goog-api-key header sends it. No SDK install is needed for this call. Change thinking_level to medium, high or xhigh for your evaluation. The snippet leaves out max_output_tokens, so the 65,536 synchronous default applies. generation_config is the object that carries thinking_level; keep that nesting when you adapt the call.

Context, output limits and tool continuation

Input, thinking and output share one 1,048,576-token window. A synchronous call stops at 65,536 output tokens. For the Interactions API, Google's long-decode guide allows up to 1,048,576 minus your input tokens, delivered in 64K-token hops. Background mode requires store: true, can run for up to 24 hours, and still stops at 65,536 if you omit max_output_tokens. The long-decode guide also covers foreground client-side continuation.

max_output_tokens caps combined thinking and output. A hard cap can stop generation with status incomplete, truncated text or no text, so lower thinking_level when you want to cut cost or latency. In the Interactions API, multi-turn tool use can keep state with store: true and previous_interaction_id. In stateless Interactions mode, resend every thought block unchanged. GenerateContent is different: history is client-managed. Return prior model turns and thought signatures unchanged, and do not copy a signature across parallel functionCall parts.

What one request costs at Google's standard rate

Google's standard paid price is $1.50 per 1M input tokens and $7.50 per 1M output tokens, with output including thinking. Cached input is $0.075 per 1M through implicit caching; Google supports implicit caching only. These are Google's list prices at the snapshot time above.

For an illustrative request with 10,000 uncached input tokens, 1,000 visible output tokens and 1,000 thinking tokens, input costs 10,000 × $1.50 / 1M = $0.015. Billed output is 2,000 tokens, so output costs 2,000 × $7.50 / 1M = $0.015. The total is $0.03. This arithmetic excludes tools, retries and cache hits.

Google's free tier is rate-limited. Batch, Flex and Priority inference are separate consumption options with their own prices; do not apply their rates to a standard interactive request. Carbon and Argon have no free tier and list $2 input and $10 output per 1M through November 30, 2026, then $4 and $20 from December 1.

Hydrogen, Carbon and Argon: positioning and migration

The three Gemini 4 models share the same 1M unified context, long-decode design and built-in tools. Google separates them by reasoning depth, latency and token use.

Google calls Carbon its strongest frontier model, built for hard architecture, debugging, science and security audits. It describes Argon as the choice for most advanced long-horizon tasks with lower token use than Carbon. Hydrogen is the fast, efficient option for everyday coding and agent execution. Google's pages do not name one winner across all tasks; choose by the work you have and test on your own prompts.

For the larger-model decision, see the Carbon vs Argon comparison. For a coding-focused Argon walkthrough, see the Argon coding test.

Switching from Carbon or Argon needs no request-schema change: set model to gemini-4-hydrogen, then set thinking_level if you want a level other than high. Hydrogen supports 65,536 tokens per non-streaming hop, while Carbon and Argon support 32,768 per non-streaming REST hop. In the Interactions API, if you switch models across turns, use previous_interaction_id or preserve thought steps in stateless mode.

From Gemini 3.8 Flash, raise or remove low max_output_tokens caps, remove explicit caching fields and sampling parameters (temperature, top_p, top_k), and preserve thought signatures. Implicit caching starts at 4,096 prompt tokens. Google's what's-new guide has the full migration list.

What to check before switching

Run this checklist before changing a production model string.

1. Use the exact ID gemini-4-hydrogen; do not guess a gemini-4-flash ID.
2. Budget input, thinking and output inside the shared 1,048,576-token window.
3. If one answer needs more than 65,536 tokens, use Interactions long decode: client-side continuation, or background mode with store: true.
4. Test medium and high first; add xhigh only for tasks where high falls short.
5. Remove explicit caching and sampling fields; rely on implicit caching only on repeated prompt prefixes of at least 4,096 tokens.
6. Choose stateful Interactions API (store: true, previous_interaction_id) or client-managed GenerateContent histories, and keep thought signatures intact.
7. Confirm the tool and modality: text, image, video, audio and PDF input with text output; no image generation, audio generation or Live API.

Frequently asked questions

Is Gemini 4 Hydrogen free?

Google offers Hydrogen on a free tier that is rate-limited, plus a paid tier at $1.50 per 1M input and $7.50 per 1M output.

Is Gemini 4 Hydrogen the same as Gemini 4 Flash?

Google lists Hydrogen as its own model; no official Google page equates it with Flash. The Gemini 4 Flash release-date page tracks the separate Flash question. Use the Hydrogen ID for Hydrogen and wait for a published Flash ID before calling Flash.

Does high mean the same compute on Hydrogen, Carbon and Argon?

No fixed token budget or controlled equal compute comes with the label. Google describes each level by relative token use, and the models have different defaults and maximum levels: Hydrogen defaults to high and supports xhigh; Carbon and Argon default to medium and stop at high. For a first comparison, use the same task, cap and explicit level, then measure total tokens and accepted results, and tune each model.

What is Hydrogen's maximum output?

Synchronous calls stop at 65,536 output tokens. In the Interactions API, long decode can raise the cap to 1,048,576 minus input tokens, in 64K-token hops with background mode or client-side continuation. Thinking tokens count toward that cap.

Sources checked

• Google: Introducing Gemini 4 Hydrogen (November 12, 2026)
• Google: Gemini 4 Hydrogen model page (checked November 13, 2026 UTC)
• OmniaKey: Gemini 4 Hydrogen model page (catalog listing, checked November 13, 2026 UTC)
• Google: What's new in Gemini 4 Hydrogen (checked November 13, 2026 UTC)
• Google: Gemini API changelog (November 12, 2026)
• Google: Gemini API pricing (checked November 13, 2026 UTC)
• Google: Thinking guide (checked November 13, 2026 UTC)
• Google: Long decode (checked November 13, 2026 UTC)

Hydrogen is now listed on OmniaKey; see the Gemini 4 Hydrogen model page for current capabilities and pricing.