Google has added a new mode to its Antigravity coding platform that lets two AI models share a single task, with one planning the work and the other writing the code. Called Orchestrator, it pairs Gemini 4 Carbon with Gemini 4 Hydrogen inside one session, handing control back and forth as the job evolves. Google says the combination reaches near-Carbon results while using 24% less quota and finishing 19% faster.
How Orchestrator works
The idea is simple: different stages of a coding task need different strengths. Carbon starts in a read-only phase, exploring dependencies, resolving ambiguities and building a concrete plan with explicit triggers for re-planning. Hydrogen then takes over to write the edits, update packages and run tests at speed.
If Hydrogen hits friction — a failing test, a compiler error — control hands back to Carbon, which inspects the state, repairs the plan and resumes execution. Google calls this a dynamic, bidirectional approach, in contrast to static routers that guess a task’s difficulty from the first prompt and send the whole job to one model.
The numbers behind it
On Terminal-Bench 4.1, Google’s benchmark for agentic coding, Orchestrator reached near-Carbon task success rates while consuming 24% less quota and completing tasks 19% faster on average. Hydrogen on its own scored 57% on the same benchmark, and 51 on the Artificial Analysis Intelligence Index, passing Argon’s 50 and reaching 54 at its highest reasoning setting.
The figures are Google’s own, measured on its own benchmark, and they describe a specific setup: long-horizon tasks in Antigravity where planning and execution can be separated. They are not a claim that two models are better than one for every job — only that, for the jobs Orchestrator targets, the pairing saves quota without giving up much accuracy.
Who gets it
Orchestrator is rolling out as Orchestrator (Carbon) and Orchestrator (Argon) for Google AI Ultra subscribers, and as Orchestrator (Argon) for paid Google AI Pro subscribers. Enterprises can access both through Gemini for business. Hydrogen itself is rolling out across paid Pro and Ultra subscribers, with four reasoning levels from low to a new xhigh tier.

Our Read
Orchestrator is Google’s answer to a problem every coding-agent maker now faces: the smartest model is too expensive to run on every step, and the fastest model is too weak for the hard ones. Splitting the work inside one session is the obvious fix, and Google’s numbers suggest it works.
The comparison that matters is not with Carbon alone but with the alternatives. Hydrogen is now the default in AI Mode and Search, and Carbon arrived two days earlier with its own multi-agent mode, /swarm. Orchestrator sits between them: less ambitious than a swarm of agents, more useful than a single model for a long task.
The catch is access. Orchestrator (Carbon) is limited to Ultra subscribers, so the developers who would benefit most — those running long tasks on tight budgets — mostly cannot use it yet. Google says more plans will follow; until they do, Orchestrator is a showcase for Ultra, not a tool for everyone.
Why It Matters
It matters because multi-model coding is becoming the norm, not the exception. Developers mix models by hand — a frontier model for planning, a fast one for edits — and the tools that automate that split will set the pace. Google’s version builds the handoff into the editor, with the planner able to repair the plan when execution goes wrong.
It also matters for Google’s subscription ladder. Orchestrator is the clearest reason yet to pay for Ultra, and its limits — Carbon for Ultra only, Argon for Pro — draw the line between the tiers more sharply than any model list. Antigravity replaced Jules in October; Orchestrator is the feature that makes that bet pay.
Background
Google launched Antigravity in May 2026 and has added specialised models for different engineering tasks since, including drug-discovery skills in November. Orchestrator arrived alongside Gemini 4 Hydrogen on 12 November, two days after Carbon, and tens of thousands of /swarm sessions had already been launched in the days before.
The Antigravity SDK also lets developers wire Hydrogen and Carbon together directly for multi-agent pipelines, with Hydrogen agents handling triage and coding while a Carbon agent takes larger refactors and security checks. Orchestrator is the no-code version of the same idea, built into the model picker.
Source: Antigravity blog.


0 Comments