What happened
Google’s Gemini API page identifies the stable model as gemini-4-hydrogen and describes it as designed for everyday coding and agentic workflows. The page gives a unified input-and-output context limit of 1,048,576 tokens. The default maximum output for synchronous requests is 65,536 tokens; with long decode, output can reach 1,048,576 tokens. That distinction matters operationally: the shared context ceiling does not mean an ordinary synchronous response can return that many tokens.
The documentation accepts text, image, video, audio and PDF inputs, while listing text as the only output type. Its capability table marks agentic video understanding, code execution, computer use, context caching, file search, function calling, grounding with Google Maps, grounding with Google Search, structured outputs, thinking and URL context as supported. It explicitly says that explicit caching is not supported, despite the broader context-caching row being marked supported. The page marks audio generation, image generation and Live API as unsupported.
Google also lists Batch API, Flex inference and Priority inference as supported consumption options. The page names the model version gemini-4-hydrogen as stable and dates its latest update to November 2026. It directs readers to a separate “What’s new in Gemini 4 Hydrogen” page for feature coverage, API changes from Gemini 3.8 Flash and migration guidance. It links to a model card and pricing page, but the model documentation page itself does not state a price.
The resulting picture is more specific than a broad claim that Hydrogen supports agentic work. A team can check whether its workflow depends on computer use, code execution, file search, structured output or a particular grounding path; it can also identify mismatches such as a need for generated media or Live API. The documentation is a compatibility reference, not a demonstration that each capability meets a given application’s quality, latency or reliability requirements.
What the archive adds
New Runtime’s earlier coverage introduced Gemini 4 Hydrogen for agentic workflows on November 12, 2026. This documentation is a new fact on that known story: it makes the model’s API limits and capability boundaries explicit, rather than announcing another model event.
Trend signal
This documentation does not yet fit a pattern New Runtime tracks. It specifies a single model’s API surface, but does not establish a broader shift in how agent interfaces, runtimes or interoperability are designed.
Why now
The page dates its latest update to November 2026 and points readers to migration guidance describing API changes from Gemini 3.8 Flash. That makes the specification relevant to teams evaluating a move between model generations: they can compare their existing workflow requirements against the stated support matrix before changing integration code. The documentation does not state a migration deadline.
Connections to other events
New Runtime’s earlier coverage also noted that Gemini 4 Hydrogen became available through Vercel AI Gateway on November 12, 2026. The AI Studio model page answers a different implementation question: which API capabilities and limits Google documents for the model itself. Gateway availability does not, by itself, establish that every model capability or limit is exposed identically through another provider.
Why it matters
A model’s advertised context window is only one part of whether it fits an agent workflow. Teams need to know whether the model can use the tools their system depends on, what kinds of input it accepts, what it returns and how much output a normal request can produce. Hydrogen’s page gives builders a concrete starting checklist: computer use and code execution are supported, but generated audio, generated images and Live API are not; default synchronous output is 65,536 tokens, even though long decode can reach 1,048,576.
That can prevent avoidable integration work. A team that needs a long text artifact can investigate long decode, while a workflow that depends on image generation should not infer support from the model’s image input capability. These are compatibility facts, not evidence that the model will perform a task accurately or cost-effectively. Builders still need to test their own workloads and verify the behavior of each API path they plan to use.
Limits and uncertainty
The documentation states feature support and token limits, but does not quantify quality, latency, reliability or cost for specific workloads. It links to a pricing page without stating prices in this model entry. The page also does not explain the distinction between general context caching and unsupported explicit caching, so teams should consult the linked caching documentation before relying on a particular caching behavior. The capability table is Google’s specification; New Runtime has not independently tested these features or confirmed how they behave in production applications.
Skill to build now
Make a short compatibility checklist before migrating an agent: record each required input type, tool call, grounding method and output format, then compare those requirements with the model’s capability table. Separately label tasks that need more than 65,536 output tokens, since they require long decode rather than assuming the standard synchronous limit will suffice. Treat unsupported entries as blockers to investigate, not as features to work around by assumption.
Reader takeaway
Gemini 4 Hydrogen’s API documentation is useful because it separates a large context window from actual workflow compatibility. Check the individual capability flags and the default output cap before adopting it; the million-token long-decode ceiling is a distinct mode, not the ordinary synchronous limit.