Gemini 4 Hydrogen: Price, Speed and Benchmarks

Google released Gemini 4 Hydrogen on November 12, 2026, and put it almost everywhere at once: Google Search, the Gemini app, the Gemini API, AI Studio, Gemini for business and Antigravity for paid AI Pro and Ultra subscribers.

The pitch is not that Hydrogen is the smartest model. Google’s frontier models, Gemini 4 Carbon and Gemini 4 Argon, keep that job. Hydrogen is the fast, cheap execution model for everyday coding and agentic workflows, and the company is positioning it as the default rather than an upgrade.

The confirmed numbers

Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1-million-token context window.
Positioning: above Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default high thinking level, 51 versus 50, reaching 54 at the new extra-high “xhigh” setting.
Speed: Google says tasks finish 2.5 times faster than Gemini 3.8 Flash at the high default setting.
Coding: Terminal-Bench 4.1 scores of 57 percent versus 18.2 percent for Gemini 3.8 Flash, per Google.
Reliability claim: fabricated claims cut by more than five times, and better behaviour when a tool call fails or a search returns nothing.
Thinking levels: low, medium, high (default) and xhigh.

Why a flat price across a million tokens matters

Many models charge more once a prompt crosses a length threshold. A flat rate across a million-token window makes long-context work predictable to budget: a 400,000-token repository dump costs the same per token as a 4,000-token question.

At $1.50 and $7.50, Hydrogen sits well below what Google charges for its frontier models, and that gap is the whole reason the model exists: it lets Google route most queries to something cheap without routing them to something weak.

The Orchestrator wrinkle

Google also introduced Orchestrator in Antigravity, which switches between models inside a single session rather than using one model start to finish. On Terminal-Bench 4.0, Google says Orchestrator reached near-Carbon task success rates while using an average of 24 percent less quota and finishing 19 percent faster.

That is a more interesting claim than any single benchmark number, because it is about the system rather than the model. It is also entirely vendor-reported.

What this changes in Search

The same model now powers a Search upgrade Google describes as its biggest model change yet: better handling of multi-part questions and typos, more aggressive query fan-out to find links, and more visual answers with comparison tables and custom tools. Google says the Search rollout is global for all users.

Practical effect for publishers and anyone who relies on search traffic: the answer surface gets richer, and the click becomes more optional. That is a trend worth watching rather than a sudden change.

What the benchmarks do not prove

Every performance figure above is Google’s own. Three limits are worth keeping in mind:

Benchmarks are chosen by the vendor. Terminal-Bench measures a specific kind of agentic coding task. A high score does not tell you how the model handles your documents, your language or your tool chain.
“Fabricated claims cut by 5x” is not “no fabrication”. A fivefold reduction from a base rate you cannot see leaves an unknown residual rate.
Default routing changes the comparison. If free Gemini app users get Hydrogen through automatic model selection, many people will use it without choosing it, and their experience will depend on routing decisions they cannot inspect.

For context on how quickly this lineup is moving, our explainer on Gemini 4 Carbon and the /swarm agent feature covers the model Google released two weeks earlier, and OpenAI’s models arriving on Google Cloud is the commercial side of the same story.

What we did not do

We did not run independent benchmarks, and we have not tested Hydrogen against a fixed task set. Prices, benchmark figures, availability and the Orchestrator numbers all come from Google’s November 12 announcement. We have not quoted a frontier-tier price we could not verify, and we have not repeated third-party rankings as fact.