Gemini 4 Hydrogen Explained

Gemini 4 Hydrogen is a fast, low-priced Google model announced on November 12, 2026, positioned as the "daily driver" for everyday coding and agentic work. Its model ID is gemini-4-hydrogen, and it costs $1.50 input / $7.50 output per 1M tokens, flat across the whole 1M-token context window. If Gemini 4 Carbon and Gemini 4 Argon are the frontier models for deep reasoning, Hydrogen is the one built for rapid execution.

Three-line summary.

- What it is: a fast successor to Gemini 3.8 Flash. 57% on Terminal-Bench 4.1 (3.8 Flash: 18.2%) and 51 on the Artificial Analysis Intelligence Index (54 at xhigh)
- What it costs: $1.50 input / $7.50 output / $0.075 cached input per 1M tokens (thinking tokens count as output). Double the 3.8 Flash introductory price ($0.75 / $3.75), but below the Argon and Carbon introductory price ($2 / $10)
- Where to use it: rolling out from launch day in the Gemini API, AI Studio, the Gemini app, Google Search and Google Antigravity, among others. Four thinking levels: low, medium, high (default) and xhigh

This column organizes what Google's official blog posts and the Gemini API pricing and model documentation say about pricing, benchmarks, availability, API usage and thinking levels. Figures are Google's own and have not been independently reproduced.

The official blog leads with speed and reliability gains over 3.8 Flash. At the default high setting it finishes tasks 2.5x faster, and its Terminal-Bench 4.1 score more than triples. In partner tests on graph, SQL and data-heavy enterprise workflows it was 2x to 2.8x faster while using 12% to 23% fewer tokens. In the Google Sheets canvas, it generates output more than 30% faster than previous models.

The second theme is behavior when run as an agent. Google says overreach in agentic tasks fell by up to 81% and under-service by up to 79%, misaligned actions and task shortcutting across tool-use and coding evals dropped 12x, and the model is up to 2x less likely to bypass permission boundaries when a task is impossible or blocked. Resilience to indirect prompt injection is more than 2.3x on the Gray Swan IPI benchmark, and chain-of-thought legibility is 99.9% with zero deceptive reasoning edits across more than 200,000 audited turns.

A new top thinking level, xhigh, has also been added. It is not available on Gemini 3.8 Flash and lets you tune cost and quality within a single model. Google Antigravity also offers an "Orchestrator" setup in which Carbon coordinates while Hydrogen powers worker and verifier agents.

According to the Gemini API pricing page, Hydrogen's standard price is $1.50 input / $7.50 output (including thinking tokens) / $0.075 cached input, and the unit price does not change anywhere in the 1M-token context. Caching is implicit only; explicit caching (such as cachedContents) is not supported. Implicit caching applies to prompts of at least 4,096 tokens.

Comparison with other models. The pricing page lists price changes on January 1, 2027 for 3.8 Flash and on December 1, 2026 for Argon and Carbon (shown as "now → after change" below).

What this shows: Hydrogen's price equals the post-2027 price of 3.8 Flash ($1.50 / $7.50). Through 2026, 3.8 Flash costs half as much, but from the new year both cost the same, at which point Hydrogen leads on speed and capability. Argon and Carbon move to $4 / $20 on December 1, roughly 2.7x Hydrogen's input price. See also our columns on Gemini 4 Carbon and Gemini 4 Argon.

A sample monthly estimate assuming 10M input tokens and 2M output tokens (standard pricing, thinking counted as output).

This simply multiplies unit prices. Real bills depend heavily on thinking-token volume and how long each model's outputs run. Heavy use of xhigh produces more output tokens, so the bill rises even at the same unit price. Grounding with Google Search is billed separately per query.

The comparison table in the official blog is an embedded image (GIF), so only figures stated in the page text are listed below. Competitor numbers that appear only inside the image could not be verified and are omitted.

Three points on reading this. First, the biggest change is the gap to 3.8 Flash: Terminal-Bench 4.1 goes from 18.2% to 57%, and fabricated claims fall by more than 5x. Second, the Intelligence Index score of 51 edges out Argon's 50, but that is one index; it does not mean Hydrogen replaces Carbon or Argon on long-horizon or difficult tasks. Third, the Orchestrator setup (Carbon coordinating, Hydrogen executing) is described on the Antigravity blog as reaching near-Carbon task success on Terminal-Bench 4.1 with about 24% less quota and about 19% faster completion on average (the main Google blog says Terminal-Bench 4.0 for the same claim, so the version label is inconsistent).

Also, the comparison images in the Search post are labeled "Gemini 3.5 Flash" while the text says "Gemini 3.8 Flash". Treat the text, not the image, as the reference for the comparison model.

For Search, Google says understanding of complex multi-part questions, typos and highly specific requests improves, query fan-out surfaces a wider range of links, and responses include more visual elements such as comparison tables. No country-by-country or language list is given (it says global and all supported languages).

The official samples use the Interactions API (client.interactions.create) with the GEMINI_API_KEY environment variable. Here is a minimal example with the Python SDK (google-genai).

With curl (the official sample, setting thinking_level to xhigh):

To produce a response longer than 65,536 tokens, set background: true and max_output_tokens. The documentation suggests 131,072 or 262,144 for long outputs. Because output spans 64K continuation hops, you must join the text across all model_output steps.

Points to check before implementing.

- Remove sampling parameters: temperature, top_p and top_k are ignored by the server, and removing them is recommended
- minimal is rejected: passing it as thinking_level returns an error
- max_output_tokens caps thinking plus output combined: long thinking leaves less room for the answer
- No explicit caching: cachedContents and caches.create are unsupported. Only implicit caching applies, for prompts of at least 4,096 tokens
- Declare tools: prompting for raw JSON tool calls without declared tools can return MALFORMED_FUNCTION_CALL. Declare tools natively or use structured outputs
- Thought signatures: do not copy a thought signature across parallel functionCall parts
- What is billed: thinking tokens, plus internal system instructions and function declarations

Hydrogen defaults to high. Carbon and Argon default to medium and cap at high, so swapping only the model name in existing code changes behavior. Non-streaming hops also allow 65,536 tokens on Hydrogen versus 32,768 on Carbon and Argon. Google says cost, latency and quality scale predictably across levels.

The "rough fit" column is this column's own framing, not an official recommendation. What the documentation says is that xhigh increases latency and token use and can overthink simple prompts, so evaluate at all levels before choosing one. Prepare 10 to 30 of your own tasks and compare low through xhigh under identical conditions.

If you work with video generation or video input, see also our column on Gemini Omni 1.2 Flash. Details on the predecessor are in Gemini 3.8 Flash.

- All figures in the official benchmark comparison image, including competitor models (only numbers stated in the text are listed)
- Knowledge cutoff date
- Vertex AI availability (not mentioned on the official pages we checked)
- Whether Hydrogen's price will change in the future (the pricing page lists no change)
- Concrete quota limits for Hydrogen alone in Antigravity
- Country-by-country and language lists for Search
- Any migration deadline or deprecation plan for Gemini 3.8 Flash

Gemini 4 Hydrogen is the execution model of the Gemini 4 generation: $1.50 / $7.50 flat across the 1M context, and a Terminal-Bench 4.1 score more than triple that of 3.8 Flash. 3.8 Flash stays cheaper through 2026, but the two cost the same from January 1, 2027. A practical start is to run your own tasks on gemini-4-hydrogen at the default high level and compare against low and xhigh. The figures are announcement values, so confirm them on your own evaluation set before production use.