Two weeks after Gemini 4 Carbon, Google has filled in the bottom of the Gemini 4 lineup. Gemini 4 Hydrogen came out on November 12, 2026. Google positions it plainly as the model that does the work: Carbon and Argon handle deep, long-horizon reasoning, and Hydrogen handles “rapid execution” (writing code, running tests, checking changes).
It’s available today in the Gemini API and AI Studio (model ID gemini-4-hydrogen), Android Studio, Gemini for business, the Gemini app, Google Search, and Google Antigravity for paid AI Pro and Ultra subscribers.
Price and limits
| Gemini 4 Hydrogen | Gemini 4 Carbon (intro price) | |
|---|---|---|
| Input / 1M tokens | $1.50 | $2.00 ($4.00 from Dec 1) |
| Output / 1M tokens (incl. thinking) | $7.50 | $10.00 ($20.00 from Dec 1) |
| Cached input / 1M | $0.075 (implicit only) | $0.10 |
| Context / max output | 1M / 1M tokens | 1M / 1M tokens |
The price is the same across the whole 1M-token window, with no surcharge for long prompts. Batch and Flex are half price ($0.75 / $3.75), and Priority costs $2.70 / $13.50, per LiteLLM’s day-0 pricing notes.
Caching works differently from Carbon. Hydrogen supports implicit caching only, with a 4,096-token minimum prefix. You can’t create a cached_content object. To get cache hits, keep the system prompt, tool definitions and reference docs at the top of the request, unchanged between calls.
Four thinking levels
Hydrogen has low, medium, high (default) and xhigh thinking levels. Google’s headline claim is that at the default level it edges out its own Argon on the Artificial Analysis Intelligence Index (51 vs 50), and reaches 54 at xhigh. Other figures Google reports against Gemini 3.8 Flash:
- Terminal-Bench 4.1: 57% vs 18.2%
- 2.5x faster task completion at the default setting
- More than 5x fewer fabricated claims
- Partners saw 2x–2.8x faster runs with 12–23% fewer tokens, partly because the model batches file reads and searches in parallel instead of reading files one at a time
All of these are vendor-reported. Treat them as a reason to run your own comparison, not as a result.
Orchestrator: routing built into the IDE
The more interesting part for agent builders is Orchestrator in Google Antigravity. Instead of picking one model per session, Orchestrator switches between Carbon and Hydrogen within the same session. On Terminal-Bench 4.0, Google says it got close to Carbon’s success rate while using 24% less quota and finishing 19% faster.
For parallel work, Google describes a familiar pattern: Carbon as coordinator, Hydrogen as the worker and verifier agents, each building and testing changes in its own isolated workspace. It’s the same planner/executor split many teams already build by hand. The difference is that the vendor now ships it as the default way to use its models.
Safety numbers worth noting
Google says Hydrogen reduces agentic “overreach” by up to 81%, is up to 2x less likely to bypass permission boundaries on blocked tasks, and is 2.3x more resilient on the Gray Swan indirect prompt injection benchmark than 3.8 Flash. These numbers matter more for a cheap model, because cheap models are the ones you run unattended in large numbers.
What to do with it
- If you route models yourself, Hydrogen is a new candidate for your executor/subagent slot. At $1.50 / $7.50 it’s well under Carbon’s post-December price, so test it against your current cheap tier on your own tasks before switching.
- Move stable context to the front of the request. With implicit-only caching, a stable prefix decides whether you pay $1.50 or $0.075 per million input tokens.
- Start at the default level, and use xhigh only on tasks where a check (tests, a verifier) shows that low or high fails.