Gemini 4 Hydrogen finally separates planning and execution
Google deploys Gemini 4 Hydrogen, a model designed for code, agents, and large-scale uses. Its working logic separates architecture and execution, with an API price starting at 1.50 dollars per million input tokens.
Google launched Gemini 4 Hydrogen on November 12, 2026, with a global rollout in its products for developers, businesses, and the general public. This model is mainly designed for fast execution: writing code, running tests, verifying changes, conducting research, and managing tool-equipped agents. Its value is concrete: it complements Gemini 4 Carbon, which is more comfortable with planning and long reasoning, rather than trying to do everything in one go.
A model to go fast, not to decide everything
Hydrogen is presented as an adapted variant of Gemini 4 Carbon, with a distinct role in the lineup. Carbon and Argon handle architectural trade-offs and deep reasoning; Hydrogen is designed for rapid execution. It accepts text, images, audio, and video, with a knowledge cutoff set to March 2026, which may make it less up-to-date in some areas.
This separation matters because it reflects a clear trend in AI models: sometimes it’s better to have one brain for planning, then another for acting. The idea is not just elegant on paper; it reduces unnecessary back-and-forth when a task alternates between exploration, code modification, and verification.
In Antigravity, the Orchestrator mode changes roles on the fly
Google Antigravity’s new Orchestrator mode applies this division within the same session. Carbon can first explore dependencies and produce a plan in read-only mode, then Hydrogen takes over to modify code, update packages, and run tests. If execution stalls, control can return to Carbon to inspect the state and adjust the strategy.
The difference with a static router is significant: instead of assigning the entire task to a single model based on the first instruction, Orchestrator can switch tools as work progresses. Google states that the Antigravity team achieved results close to Carbon on Terminal-Bench 4.1, with 24% less quota and an average execution time reduced by 19%.
For parallel uses, Carbon can coordinate while Hydrogen powers worker and verifier agents in isolated spaces. Hydrogen keeps a low profile: it executes, then steps aside.
What Google says about performance and limits
Figures presented by Google show a model faster than Gemini 3.8 Flash on certain scenarios. At a high reasoning level, Hydrogen completes tasks 2.5 times faster, scores 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash, and reduces fabricated assertions by more than five times. Its Artificial Analysis Intelligence Index score reaches 51 in high mode, versus 50 for Argon, and climbs to 54 in xhigh.
But the comparison is not uniform. The model sheet also indicates that Claude Sonnet 5.5 scores 63.9% on an adjusted version of Terminal-Bench 4.1, ahead of Hydrogen’s 57.0% [to be verified]. In other words, Google significantly improves its previous Flash model without dominating the entire dashboard.
Limits are real too. Documentation notes possible hallucinations, occasional slowdowns or timeouts, and higher token consumption as reasoning level increases. In AI, thinking longer always costs a bit more; the bill, however, does not read footnotes.
API pricing, context window, and reasoning settings
Hydrogen’s API pricing is set at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens. Google adds a free tier with rate limiting, as well as an implicit cache charged at 0.075 dollars per million input tokens. The total context window reaches 1,048,576 tokens, with synchronous responses limited to 65,536 tokens per round trip.
The model offers four reasoning levels: low, medium, high by default, and xhigh. Google recommends high for most complex tasks and reserves xhigh for more difficult multi-step problems, noting that this setting may overthink simple instructions. This choice directly affects latency and available space in the context window.
Wide deployment in app, Search, and paid offerings
The November 12, 2026 deployment covers Gemini for businesses, the Gemini app, Google Search, Antigravity, and Gemini in Sheets for eligible subscriptions. Free Gemini app users will receive Hydrogen via automatic model selection when Google deems the response relevant. Google AI Pro and Ultra subscribers will also have access in Antigravity, while the company promises access via Gemini for business.
Google also claims the app better follows tone and format instructions, relies more faithfully on uploaded files and connected services like Gmail, Drive, Airtable, and Dropbox, and chooses its tools more reliably. Search receives improved handling of multiple or misspelled queries, more visual responses, and expanded search for relevant links.
A real-world test for agent models
The real question is not only whether Hydrogen codes faster. It is whether this division between planning and execution will hold in real uses, with noisy tasks, connected services, and reliability constraints. Google claims efficiency gains, but its own documents also mention unjustified refusals increasing by 5.57 points and security still incompletely documented.
Hydrogen thus shows a clear direction: reserve heavy reasoning for the steps that need it, and let a faster model run the rest. If the balance holds, it is a more sober way to orchestrate agents. Otherwise, we will mostly have gained a new way to argue with our bugs.
Key points
- Global launch announced for November 12, 2026.
- Hydrogen targets code, tests, and agents.
- Google cites 24% less quota on Antigravity.
- The xhigh mode complements reasoning settings.
- Claude Sonnet 5.5 is given at 63.9% [to be verified].
- Access planned for Gemini app, Search, and Google AI Pro.
By the numbers
- 1.50 dollars — million input tokens, API price announced by Google.
- 7.50 dollars — million output tokens, API price announced by Google.
- 1,048,576 tokens — total context window, Gemini API documentation.
- 57% — Terminal-Bench 4.1 score, Google model sheet.
- 79% — reduction of agent under-service compared to Gemini 3.8 Flash.