Gemini 4 Hydrogen: API Pricing and Everyday Agent Work

The stable Gemini Developer API model ID is gemini-4-hydrogen. AI Studio offers an interactive way to try a prompt before integrating it into an application. The model is available on free and paid API tiers; their limits and data-use terms differ. Paid-tier data is not used to improve Google products, while the free tier permits that use.

Antigravity is a coding environment with its own tools and subscription quota. At launch, Hydrogen is available in Antigravity to paid Google AI Pro and Ultra subscribers. Gemini app and Search access are separate product routes: free app users can receive Hydrogen through automatic model selection. An app subscription does not pay your separate Developer API bill.

Cursor also offers Hydrogen in its model picker and recommends high effort for agentic coding. Its model page lists 200K default context and up to 1M maximum context. Hydrogen uses Cursor’s third-party model billing: prepaid credit on The X Subscription, or the Other Models pool on legacy plans. Use Cursor’s allowance and usage records when budgeting that route.

Set GEMINI_API_KEY in your local shell using a key for the intended Google project. The following REST request uses the Interactions API and an explicit medium thinking level. Read the returned text and usage, then run any generated code in your own development environment. The request returns code; it does not grant access to your repository or run a test suite.

For a repository agent, connect the model through a client that supplies file, shell and test tools. Give it one issue and a concrete acceptance test. Include the tool results in subsequent turns and inspect the final diff. Tool declarations, repeated source files, retries and reasoning all affect the complete session’s usage.

Checked November 13, 2026, these are Gemini Developer API USD rates per million tokens. The standard input/output rate remains flat across the supported context window. Output includes billed thinking tokens. Cached input uses implicit caching only; explicit cache creation is unsupported. Batch and Flex cost less per token but serve different latency and scheduling needs.

For an illustrative task totaling 100,000 fresh input tokens and 20,000 billed output tokens across all turns, Standard costs $0.15 + $0.15 = $0.30. The same token totals cost $0.15 on Batch or Flex and $0.54 on Priority. These totals exclude paid tool services and assume every input token is fresh.

If 80,000 of those input tokens qualify for Standard cache-hit pricing, the example becomes $0.03 fresh input + $0.006 cached input + $0.15 output = $0.186. Use the response usage breakdown to apply the cache rate, rather than assuming that every repeated prompt is a cache hit.

High is the default. Google recommends it for complex coding and multi-step work. Medium targets interactive coding and latency-sensitive tool calls; low minimizes reasoning overhead for simpler work. Xhigh spends more tokens and time on difficult problems and can overthink a simple prompt.

Compare levels on a task with a checkable result. A short response that needs three repairs can cost more than one successful high-effort attempt. Record total input, billed output, tool calls and elapsed time, then choose the lowest level that reaches your acceptance criterion. The max_output_tokens setting bounds thinking plus final output, so an overly small cap can truncate the answer.

The documented unified window is 1,048,576 tokens, shared across input, thinking and output. Synchronous responses allow up to 65,536 tokens per hop. Longer generation uses background execution or continuation, within the remaining shared window. Loading almost a million tokens of source material leaves less room for reasoning and the result.

Hydrogen accepts text, images, video, audio and PDFs and returns text. Function calling, structured output, code execution and computer use are supported; image generation, audio generation and the Live API are not. Choose a generative media model when the output itself must be an image or speech recording.

When migrating from Gemini 3.8 Flash, check the combined context budget and output limits as well as the model name. Preserve the API’s required conversation state when using tools, and test the complete client workflow rather than substituting an isolated text response for an agent run.

Orchestrator in Antigravity switches between Carbon and Hydrogen within a session. Carbon can plan or investigate a difficult blocker, while Hydrogen implements and verifies changes. This is useful when a task alternates between deep reasoning and routine execution. A direct Hydrogen API call does not automatically enable Antigravity’s orchestration. At launch, Ultra subscribers can choose Orchestrator (Carbon) or Orchestrator (Argon); paid Pro subscribers get Orchestrator (Argon).

Google reports near-Carbon task success in its Terminal-Bench evaluation with 24% less quota and 19% faster completion on average. Those are publisher evaluation results for the combined workflow. Quota savings describe Antigravity’s allowance, not a measured percentage reduction in your API invoice.

Start with Hydrogen for rapid edits, debugging and repeated tool calls. Compare Carbon for difficult engineering or scientific reasoning, and Argon for frontier reasoning with lower token spend per turn than Carbon. Antigravity Orchestrator can combine these roles; direct API requests let you choose the model for each step.

For the same 100,000 fresh input tokens and 20,000 billed output tokens, Hydrogen Standard costs $0.30. Carbon or Argon costs $0.40 through November 30, 2026 and $0.80 from December 1. Hydrogen is therefore 25% cheaper at equal token volumes during the introductory period and 62.5% cheaper afterward. This calculation excludes cache and tools; different reasoning and retry usage can change the task total.

Artificial Analysis results read November 14, 2026 list Hydrogen high at 51 on its Intelligence Index and 8.3 minutes per task, against Argon high at 50 and 29.2 minutes. Hydrogen xhigh reaches 54 at 14.5 minutes; medium scores 41 and low 27. The index combines several task types, so the one-point high-effort lead is a reason for a representative trial rather than a prediction for every coding task.

Google reports 57% on Terminal-Bench 4.1 against 18.2% for Gemini 3.8 Flash, alongside 2.5 times faster completion at Hydrogen’s default high setting. This standalone comparison evaluates Hydrogen itself; the Orchestrator result above evaluates a multi-model workflow. Neither result predicts the runtime of a particular repository task.

Use the benchmark as a reason to test the model on your workflow. Hold the repository revision, tool permissions, client and acceptance tests constant, and compare completed tasks alongside total spend. Deep architectural decisions can still justify Carbon or Argon even when Hydrogen is a better fit for rapid edits.