Gemini 4 Hydrogen
Everyday repository changes, test-driven debugging, interactive web development and worker tasks with a clear goal. Use Hydrogen for repeated edit, run and refine cycles.
Turn an issue, a design brief or a collection of documents into code changes, interactive interfaces and structured findings. Gemini 4 Hydrogen is built for everyday coding and agentic execution, with a 1M-token shared context and four thinking levels for matching reasoning depth to the task.
Chat with Gemini 4 Hydrogen here. Every message is a real request, billed to your balance at the same price as the API.
Usage-based pricing with separate rates for input, cached input and output. Thinking tokens are included in output usage; a short visible answer can still use a larger output budget.
| Model | Specifications and pricing | beatapi |
|---|---|---|
| Gemini 4 Hydrogen | Gemini 4 HydrogenInput · per 1M tokens | $0.45 / 1M tokens |
| Gemini 4 Hydrogen | Gemini 4 HydrogenCached input · per 1M tokens | $0.0225 / 1M tokens |
| Gemini 4 Hydrogen | Gemini 4 HydrogenOutput · per 1M tokens | $2.25 / 1M tokens |
Choose by the work at each stage, not just the size of the prompt. Hydrogen focuses on rapid execution; Carbon and Argon handle deeper, longer-horizon reasoning. A coding workflow can use different models for planning, implementation and review.
Everyday repository changes, test-driven debugging, interactive web development and worker tasks with a clear goal. Use Hydrogen for repeated edit, run and refine cycles.
Difficult architectural decisions, subtle cross-module bugs and investigations that require exploring competing explanations before making a change.
Gemini 4 CarbonAdvanced multi-step tasks that need more sustained reasoning, with lower thinking-token use than Carbon. Keep Argon as an option between execution-focused Hydrogen and deeper Carbon investigations.
Gemini 4 ArgonHydrogen works on the execution side of an agent workflow: inspect the material, propose a focused change, use tool feedback and continue toward the requested result.
Work from an issue description, relevant files and failing tests to locate the affected behavior and prepare targeted edits. Hydrogen can follow dependencies across a repository, explain the change and write regression coverage without folding unrelated cleanup into the task.
Turn a design brief, screenshot or existing component into working frontend code. Continue with layout feedback and runtime errors to refine forms, navigation and interactive states rather than stopping at a static mockup.
Use file search, test output and application tools as feedback during a task. Hydrogen can group independent read or search requests and adjust its next action when a tool fails or returns no results. Your application supplies and executes the tools.
The model understands text, images, audio, video and PDFs, then responds with text. It can combine a report with its charts, locate details in a recording and organize the material into findings, tables or a source-linked brief.
| Benchmark | Gemini 4 Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweep | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1 (Cheating-adjusted) | 57.0% | 63.9% | — |
| Agents' Last Exam v1.1 (Score) | 71.6% | 60.4% | 69.8% |
| AutomationBench-AA | 76.1% | 71.8% | 60.2% |
| Vending-Bench 2 (Bank balance) | $14,549 | $11,165 | $9,107 |
| OSWorld-2.1 (Offline subset) | 77.6% | 83.9% | 74.2% |
| MKA (Climb WSR accuracy) | 53.4% | 52.0% | 43.8% |
| Chartography | 66.6% | 61.6% | 63.2% |
| Terminal-Bench-Science 0.1 | 51.4% | 59.9% | 58.3% |
| HLE-Diamond | 58.8% | 53.3% | 47.3% |
November 2026 evaluation: Hydrogen scores 13.5% on SWE-sweep, 57.0% on cheating-adjusted Terminal-Bench 4.1 and 53.4% on MKA. These measure different tasks: repository repair, terminal-based execution and visual knowledge work. Terminal-Bench uses high thinking and averages five attempts per task; other settings vary by benchmark. The table also shows where competing models score higher.
Combine reproduction steps, a failing test and the relevant modules. Produce a focused patch, regression tests and a change summary for a pull request. A test runner can feed failures back into the same task for another revision.
Start with product content, a visual reference and the existing frontend components. Generate a responsive implementation, then use screenshots and browser feedback to refine product selection, mobile navigation and checkout form states.
Combine report text with charts and tables to extract changes, conflicting figures and open questions. Produce a structured briefing whose conclusions point back to the supplied evidence, ready for an analyst or reviewer.
Have a planning model define interfaces, affected files and acceptance tests, then assign Hydrogen the implementation or verification stages. Your application coordinates workspaces, tool permissions and handoffs so each worker operates within an explicit scope.
Use your server-side BeatAPI key and the existing Chat Completions interface. Handle task orchestration, tool execution and responses in your application.
Sign up or sign in to open the Dashboard. Check your balance and open the credits page if you need to add funds.
Open DashboardHydrogen suits work that advances through short execution loops while still needing reasoning over code, tool results and visual material.
Use an execution-focused model for issue fixes, small features and repeated test cycles instead of spending deep-reasoning effort on every turn. Move architectural roadblocks to Carbon when the task needs a different level of investigation.
The model offers low, medium, high and xhigh, with high as its default. Use medium for interactive iteration and reserve xhigh for difficult multi-step work. Higher settings can consume more tokens and take longer; they are not a universal upgrade.
A shared 1,048,576-token window lets the model work with code, specifications and supporting evidence together. Input, thinking, tool calls and the answer share that budget, so long inputs still need space for the result.
Move between repository work, interface refinement and analysis of charts or recordings without treating every visual input as a separate task. Hydrogen returns text, which your application can turn into a patch, a report or the next tool request.
Gemini 4 Hydrogen is Google's Gemini 4 model for everyday coding and fast agentic execution, released on November 12, 2026 and adapted from Carbon. It supports repository debugging, web development, tool-using workflows and multimodal understanding. BeatAPI provides access through an OpenAI-compatible Chat Completions API for adding Hydrogen to your own application.
Hydrogen emphasizes execution and rapid iteration, while Carbon and Argon focus on deeper, longer-horizon reasoning. Use Hydrogen for well-scoped code changes and tool-driven tasks, Carbon for difficult architecture and investigations, and Argon for advanced work with lower thinking-token use than Carbon. Hydrogen also adds xhigh and defaults to high; Carbon and Argon default to medium.
No. The 1,048,576-token window is shared by input and output, including thinking, function calls and the final answer. The synchronous output limit is 65,536 tokens. Google's longer-output mode uses background execution or continuation across segments; it is not a single 1M-token response. Plan input size and answer length together.
Start with high for complex coding and tool-using tasks, medium for interactive development, and low when latency and token use matter more. Use xhigh when high does not resolve a difficult multi-step problem. Compare quality, elapsed time and total tokens on your own tasks. Raising the thinking level can increase cost and delay without improving a simple answer. These are the model's native levels; check the BeatAPI reference for supported controls.
No. Hydrogen returns text, not images, video or speech. The model can understand text, images, audio, video and PDF; BeatAPI currently provides text access. Follow the API reference for accepted inputs rather than sending native file formats to Chat Completions. Choose a dedicated image, video or speech model when your application needs media output.
Hydrogen uses a unified context budget instead of separate input and output budgets. Its default thinking level changes from medium to high, and minimal is not supported. In Google's native API, remove thinking_budget and explicit cached_content, and do not rely on temperature, top_p or top_k, which are ignored. Preserve conversation state and thought signatures in tool loops. Recheck the target API's parameter mapping before switching production traffic.
A model does not gain access to your files, terminal or accounts just because you call it. An agent application must expose approved tools, execute model-requested operations and return the results. This is how repository search, test execution and code editing become part of a continuing task. Keep write permissions and destructive actions under the application's control.
No. Orchestrator is a separate Google Antigravity feature that switches models during a task. Calling the Hydrogen model does not provide that product or automatically add Carbon. You can implement your own workflow in which Carbon plans and Hydrogen implements, but routing, workspace isolation and handoffs belong to your application.
Thinking tokens count toward output usage, so a short visible answer can still consume many output tokens. Cached input uses the separate cached-input rate in the table. Keep reusable material at the beginning and inspect usage.prompt_tokens_details.cached_tokens rather than assuming a hit. Google's native model uses implicit caching instead of explicit cache creation, with a 4,096-token minimum for eligible prompts.
Not in an ordinary text request. Web research needs a connected search or page-reading tool, or material retrieved by your application. The model's knowledge cutoff is March 2026, so recent facts need fresh evidence. Google Search grounding and URL context are native Google capabilities, not automatic features of every third-party API integration.
Match the benchmark to your task and check its setup. SWE-sweep measures repository repair, Terminal-Bench measures execution in a terminal environment, and MKA measures multimodal knowledge work. Scores depend on thinking level, tools, harness and evaluation rules. Hydrogen's 57.0% Terminal-Bench 4.1 result cannot be directly compared with Carbon's 65.2% Terminal-Bench 4.0 result because they use different benchmark versions.
Create a BeatAPI account, add credit and create an API key. Keep the OpenAI SDK, set base_url to https://direct.beatapi.io/v1 and model to gemini-4-hydrogen, then send a Chat Completions request. The API tab contains the request reference. Google subscriptions and its free API tier are separate from BeatAPI billing.
Connect Gemini 4 Hydrogen with other production-ready APIs in the BeatAPI catalog.
Guides and practical notes to help you get more from Gemini 4 Hydrogen.