Google introduced Gemini 4 Hydrogen on 12 November 2026 alongside Orchestrator, an option in its Antigravity software-development environment that switches between models during a task. Hydrogen is built for rapid coding and agent work, while Orchestrator lets a session draw on both that execution model and Gemini 4 Carbon for deeper reasoning.
Key points
- Orchestrator changes models as work progresses within a session, initially for paid AI Ultra subscribers.
- Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens through the Gemini API.
- Google’s Antigravity team reports near-Carbon task success on Terminal-Bench 4.1 with 24% less quota and 19% faster completion on average.
Orchestrator moves between Carbon and Hydrogen
Carbon handles the planning stage in Google’s division of work, while Hydrogen takes on code changes, tests and verification. Orchestrator can pass control back to Carbon if execution runs into trouble, rather than selecting a model from the opening request and keeping it for the entire task. Hydrogen is adapted from Carbon, Unite.AI reported, and also accepts text, images, audio and video.
Fixing a bug could move from a plan to code changes and tests within one session, then return to planning if execution ran into trouble. That sequence would put the choice of model alongside the work as it changes, rather than making the choice for the whole task at the start.
Hydrogen is rolling out in Antigravity to paid Google AI Pro and Ultra subscribers, while Orchestrator initially reaches the Ultra tier. Enterprise customers can access Hydrogen and Orchestrator through Gemini for business. Google had opened Carbon to wider use and previewed agent swarms, while an earlier Argon release brought that model to paying subscribers.
Antigravity reports lower quota use on Terminal-Bench 4.1
Google’s Antigravity team reports that its mixed-model approach achieved near-Carbon task success on Terminal-Bench 4.1 while consuming 24% less quota and completing work 19% faster on average, Unite.AI reported. Those are results from the team’s benchmark runs. For parallel tasks, Carbon can coordinate work while Hydrogen runs worker and verifier agents in separate workspaces.
In a separate coding comparison, a cheating-adjusted Terminal-Bench 4.1 table in Google’s model card gives Hydrogen 57.0% and Claude Sonnet 5.5 63.9%, according to Unite.AI. Google’s reported Artificial Analysis Intelligence Index result puts Hydrogen at 51 using its default high thinking setting, against 50 for Argon, and at 54 using the more intensive xhigh setting. The setting matters to both the amount of computation used and how long a response takes: Google recommends high for most complex tasks and warns that xhigh can overthink simpler prompts.
Hydrogen shares a 1,048,576-token window
Hydrogen’s input, intermediate thinking and output occupy a shared 1,048,576-token context window. Think of it as a workbench holding the material brought to a job, the working notes and the finished piece: more space taken by one leaves less for the others. Synchronous responses support up to 65,536 tokens in a single step, while background execution or client-side continuation can produce more within the shared window.
Google prices Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens across its context window. For developers building agents, the shared capacity means longer inputs and more intensive thinking draw on the same budget of available context, while the separate input and output prices apply to what the model receives and generates. Google is making Hydrogen available through AI Studio and the Gemini API, as well as Gemini for business.
Google reports mixed results in Hydrogen safety tests
Google reports reductions of up to 81% in agentic overreach and 79% in under-service against Gemini 3.8 Flash. Those measures concern failures when an agent takes actions beyond a request or falls short of carrying it out, issues that matter when a model chooses tools and acts through connected services. The model card’s automated comparisons also record a 5.57-percentage-point increase in unjustified refusals against Gemini 3.8 Flash.
Hydrogen is also rolling out to the Gemini app and Google Search. Google says free app users will encounter it through automatic model selection when the company considers it suitable for a request. The model card identifies fabricated answers, responses that sometimes take too long or fail to complete, and the possibility of greater token consumption at more intensive thinking settings among Hydrogen’s limitations.
