Gemini · Model guide
Gemini 4 Hydrogen
Gemini 4 Hydrogen is Google’s fast Gemini 4 model for everyday coding and agentic workflows, adapted from Gemini 4 Carbon. It works on its own or as a sub-agent alongside Carbon and Argon, and offers four thinking levels from low to xhigh, defaulting to high. It accepts text, images, audio, and video and returns text within a shared 1M-token input and output window.
What this model is good at
What the vendor positions it for, and which of our models to reach for instead.
VendorGoogle positions Hydrogen for rapid, disciplined execution across everyday coding and agentic workflows, while Carbon and Argon handle deep, long-horizon reasoning on the hardest problems.source
Everyday agentic coding · Google recommends it for iterative repository coding and debugging, tool-using agents, and parallel sub-agent execution.(vendor claim)source
Web development and prototyping · Google lists web development, interactive web apps, and everyday prototyping and iteration among its intended uses.(vendor claim)source
Multimodal document workflows · Google highlights more accurate reading and synthesis of tables, charts, dense PDFs, images, audio, and video.(vendor claim)source
Everyday writing and research · Google lists multimodal knowledge work and ready-to-go writing, research, and everyday work among its intended uses.(vendor claim)source
- gemini/gemini-4-carbonstronger —Choose Gemini 4 Carbon for hard architecture, debugging, science, and security work, where Google says it spends more thinking per turn to test hypotheses and verify edge cases.(vendor claim)source
- gemini/gemini-3.8-flashpredecessor —Keep Gemini 3.8 Flash when an existing integration has been validated on it and you want to compare before migrating; Google recommends Hydrogen as its upgrade.(vendor claim)source
- gemini/gemini-3.5-flash-litefaster —Choose Gemini 3.5 Flash-Lite for high-volume classification, routing, and sub-second latency targets.(vendor claim)source
Pricing and billing
Billed per token. Cached input is charged at the cache-read rate.
| Price / 1M tokens | You pay |
|---|---|
| Input | $ 1.50 |
| Output | $ 7.50 |
| Cache read | $ 0.075 |
| Cache write | $ 1.50 |
+ 2.0K × $ 7.50
Capabilities and limits
What CrossModel guarantees across every route this model can take right now.
- Context window
- 1.0M tokens
- Max output
- 65.5K tokens
- Input / output modalities
- Text + Image + Audio + Video → Text
- Streaming
- Supported
- Tool calling
- Supported
- Structured output (JSON)
- Supported
- Image input
- Supported
- Thinking mode
- Supported· always on
- Reasoning effort
- low · medium · high
- Available endpoints
- /v1/chat/completions · /v1/responses · /v1/messages
Published benchmarks
Scores the vendor reported, with the evaluation setup each one came from.
- Terminal-Bench 4.1pass_rate·Google DeepMind · 2026-11-12effort highmini-swe-agent (vision)5 runssetup not fully disclosed
Google self-computed with a mini-swe-agent vision harness at the high thinking level, average reward over 5 attempts per task, cheating-adjusted. Same table: Claude Sonnet 5.5 63.9% (public leaderboard). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
57%
- Agents' Last Exam v1.1score·Google DeepMind · 2026-11-12effort xhighAntigravity-CLI (agy) 1.4.8setup not fully disclosed
Google self-computed averaged rubric score on the 122-task v1.1 set (task version of 2026-10-17, without the 2026-10-26 update) with the Antigravity-CLI (agy) 1.4.8 harness, five-hour window, safety filters on (flagged responses return empty strings and the episode continues). Argon and Carbon were run with ALE-Claw instead. Same table: Claude Sonnet 5.5 60.4% (ALE-Claw, same setup), GPT-6.1 Terra 69.8% (public leaderboard). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
71.6%
- SWE-sweepscore·Google DeepMind · 2026-11-12effort xhighAntigravity-CLI (agy) 1.4.8setup not fully disclosed
Google self-computed with the Antigravity-CLI (agy) 1.4.8 harness, safety filters on; reports the subtask-weighted reward. Highest thinking setting, single attempt. Different harness from the Argon and Carbon results (mini-swe-agent), so not directly comparable with them. Same table: Claude Sonnet 5.5 3.1%, GPT-6.1 Terra 5.8% (public leaderboard). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
13.5%
- OSWorld 2.1 (offline subset, partial score)score·Google DeepMind · 2026-11-12effort xhighGemini CUA3 runsCompaction following Claude Opus 5.5 system cardsetup not fully disclosed
Google self-computed partial score on the 82-task offline (no-internet) subset, maximum over three single-attempt runs. Official OSWorld docker and evaluator, 1080p, max 500 steps; updated Gemini CUA harness with parallel batch tool calls, all screenshots kept in history, PyAutoGUI actuation with screenshot-only observation, safety filters on. Same table: Claude Sonnet 5.5 83.9% (Google self-computed), GPT-6.1 Terra 74.2% (its system card). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
77.6%
- Chartography (no tools)score·Google DeepMind · 2026-11-12effort xhighsetup not fully disclosed
Google self-computed without tools following the official Surge public leaderboard; highest thinking setting. Same table: Claude Sonnet 5.5 61.6% (its system card), GPT-6.1 Terra 63.2% (Google self-computed). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
66.6%
- Terminal-Bench-Science 0.1accuracy·Google DeepMind · 2026-11-12effort xhighmini-swe-agent (vision)5 runssetup not fully disclosed
Google self-computed with a mini-swe-agent vision harness, average reward over 5 attempts per task, highest thinking setting. Argon's 57.6% used a 6x verifier timeout and is not directly comparable. Same table: Claude Sonnet 5.5 59.9%, GPT-6.1 Terra 58.3% (public leaderboard). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-hydrogen
51.4%
Vendor-reported numbers, not CrossModel measurements. Scores are only comparable when the benchmark version, metric and evaluation setup match, so nothing here is averaged or ranked.
Use it in your tools
Point the base URL at CrossModel and paste this model ID — every tool below has a setup guide.
Frequently asked questions
What is Gemini 4 Hydrogen?
How much does Gemini 4 Hydrogen cost?
Does Gemini 4 Hydrogen support tool calling and structured output?
Which endpoint do I call?
Can I try it without writing code?
Chat first, integrate later
Chat first, then wire it in once you like the answers.