Google’s November 12 announcement of Gemini 4 Hydrogen, updated November 13, introduces both a model for everyday coding work and a way to switch models during a task. For artificial intelligence (AI) certification learners, it is a useful case for separating three questions: which model answers, which system assigns the work, and what it costs to finish correctly.
A model and the system choosing it
Google describes Hydrogen as a fast execution model that can work alone or alongside Carbon and Argon, its models for harder reasoning tasks. In the same announcement, Google introduces Orchestrator in Antigravity, its coding environment. Orchestrator can switch between models such as Carbon and Hydrogen within one session. The model is therefore not the same thing as the mechanism choosing a model for a particular step.
The announcement also describes a parallel arrangement: Carbon coordinates while Hydrogen worker and verifier agents build and test changes in isolated workspaces. Read that as a description of an arrangement the vendor supports, not evidence that every task benefits from adding agents. When studying an architecture diagram, label the component that assigns work separately from the components that generate code and check results.
A token price is only one input
Google lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across its one-million-token context window. Tokens are the small units used to measure the data a model processes or generates; the context window is the amount it can take into account in a request. These are published model prices, not a promise that any coding task has a fixed bill.
For a learning exercise using those listed rates, estimate the input charge and output charge separately, then add them. If the imagined workflow needs another attempt or sends work to a different model, account for that work separately too. This is a way to reason about the cost of the whole task rather than comparing only the price of one unit of output. It is not a quotation for an Antigravity subscription, a claim about available discounts or a live billing measurement.
Keep the evaluation conditions attached
On Terminal-Bench 4.0, a test of agents completing tasks in a command-line environment, Google reports that Orchestrator reached near-Carbon success rates while consuming 24% less quota and finishing 19% faster on average. Those figures belong to Google’s reported evaluation. They do not establish the same savings for an unrelated application, and quota consumption is not automatically a dollar bill.
Google also offers four thinking settings for Hydrogen: low, medium, high by default, and xhigh. The announcement presents them as choices across response quality, waiting time and cost. If you compare two configurations in a study exercise, record the setting as well as the model name. Otherwise a result attributed to the model could actually reflect different settings or a different workflow.
Design a small comparison before choosing
Here is a paper exercise, not a claim that we ran these products. Imagine a coding task that reads a small repository, changes one function and adds a test. Compare a single-model plan with a plan that uses one model to coordinate and another to implement and check the change.
- Write the acceptance conditions first: the requested behavior works, the test detects the original problem and unrelated behavior remains unchanged.
- List each model call, its thinking setting and any repeated attempt. Distinguish code that was proposed from tests that actually ran.
- Record elapsed time and estimated model charges separately. Keep a failed attempt in the record rather than discarding it because a later attempt passed.
- Explain whether any lower cost still delivers the evidence needed to accept the task. A quick answer that misses an acceptance condition is not a completed task in this exercise.
The lesson is an architecture and evaluation habit, not a requirement to use a particular vendor for an exam. Use the dated launch facts to practice identifying the model, the routing mechanism and the evidence for a completed result. Keep those three labels distinct when you assess the next announcement.
Browse the Nex Arc study guides