Google rolls out Gemini 4 Hydrogen for coding and everyday tasks
Learn where Google Gemini 4 Hydrogen is available, what it costs and how its benchmark results change with reasoning settings.
Google began rolling out Gemini 4 Hydrogen on Nov. 12, offering developers an AI model for coding and agentic workflows at $1.50 per million input tokens and $7.50 per million output tokens. The announcement appeared the same day as the Hacker News discussion, not before it.
The release gives developers a lower-priced Gemini 4 option designed to work either independently or alongside Google's more powerful Carbon and Argon models. Google's benchmarks suggest a substantial improvement over its previous Flash model, while independent tests show that Hydrogen's cost and performance vary considerably with how much reasoning it is configured to use.
Where it is available
Google says Hydrogen is rolling out globally through the Gemini API, Google AI Studio and Android Studio. It is also coming to Gemini for business, the Gemini app and Google Search. In Google's Antigravity coding environment, individual access requires a paid Google AI Pro or Ultra subscription; eligible business subscribers can also use it in Gemini in Google Sheets.
Access is not identical across those products. For example, a participant in the Hacker News discussion said free Gemini app users see an automatic selection rather than an unconditional choice of Hydrogen; that is an individual user's report, not a measured rollout status. Google describes the launch as a rollout and does not specify a general-availability designation in its announcement.
The model has four configurable thinking levels: low, medium, high and xhigh. High is the default. These settings change how much work the model can devote to a response, creating a tradeoff between quality, time and output-token consumption. Google says the API's input price is flat across the model's one-million-token context window, and that the model can produce up to one million output tokens. Those token limits do not mean every request will use that much capacity.
A faster worker alongside frontier models
Google positions Hydrogen as a successor for teams using Gemini 3.8 Flash, rather than as a replacement for Carbon on every difficult task. It can read and edit code, call tools, and work with documents, charts, audio and video. In a multi-model workflow, Google's suggested division of labor has Carbon coordinate while Hydrogen-powered agents build or verify changes in separate workspaces.
Google also introduced Orchestrator in Antigravity, which switches models within a session. In the company's Terminal-Bench 4.0 test, it says the combination achieved task success close to Carbon alone while using 24% less quota and finishing 19% faster. The figures describe Google's benchmark setup, not a guaranteed saving for every project.
For Hydrogen on its own, Google reports a 57% score on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash, and says Hydrogen finished tasks 2.5 times faster at its default high setting. The terminal benchmark tests agentic work in a command-line environment; the separately cited Orchestrator result uses version 4.0, so the two sets of percentages should not be compared directly. Google's announcement does not provide enough detail to establish how its speed and quota results will translate to a particular workload.
Google also reports 53.4% on its Multimodal Knowledge Agent benchmark, which tests agentic work over complex visual files, and 71.6% on Agents’ Last Exam v1.1, a professional-workflow evaluation. These are task-specific results, not a single measure of reliability across all documents or use cases.
What independent tests add
On the Artificial Analysis Intelligence Index, Google cites Hydrogen at 51 with high thinking and 54 at xhigh, versus 50 for Argon. The result is notable for a model Google markets as the faster execution option, though an aggregate score does not establish superiority on every task.
The Decoder's report on Artificial Analysis testing gives the tradeoff more detail: xhigh scored 54 at an average $3.16 per task, compared with 51 and $1.87 at high. Medium scored 41 at $0.80, and low scored 27 at $0.29. The outlet reports that xhigh used nearly twice as many output tokens as high, explaining much of the cost difference; those are benchmark averages, not fixed prices per customer task.
The independent terminal results also temper the broader index comparison. The Decoder reports that Hydrogen scored 54% on Artificial Analysis's Terminal-Bench 4.0 evaluation, below Claude Sonnet 5.5 at maximum thinking, at 64%, and GPT-6.1 Sol at maximum thinking, at 56%. Different benchmark versions and model settings matter: these figures should not be treated as direct confirmation or refutation of Google's Terminal-Bench 4.1 comparison with Flash.
Comments ()