1. Google Antigravity adds Orchestrator for AI Ultra
Google on November 12 began offering Orchestrator, a new mode in its Google Antigravity AI coding platform, to Google AI Ultra subscribers. Orchestrator assigns planning and review to the Gemini 4 Carbon model and hands code writing and testing to the faster Gemini 4 Hydrogen, which Google says cuts quota consumption by about 24% and task completion time by about 19%.
The idea behind the feature is straightforward. The more capable and more expensive model does the thinking that benefits most from depth, such as breaking a problem into steps and checking the results, while the quicker model handles the high-volume, repetitive work of editing files and running tests. Developers do not have to decide when to switch between them. Orchestrator appears as an option in the model selector, and once chosen, it moves work between the two models automatically within a single task.
The release arrives during a busy week for Google's Gemini 4 family. Gemini 4 Carbon itself was introduced only days earlier, and Hydrogen is now reaching users of the platform. Orchestrator is the first feature that deliberately combines the two inside one session.
2. What is Google Antigravity?
Google Antigravity is Google's AI-powered software development platform, built around agents that can plan, write, edit and test code on a developer's behalf. Rather than acting as a simple autocomplete tool, it is designed to take on multi-step engineering tasks, which is why decisions about which model handles which step have a direct effect on cost and speed.
Platforms of this kind typically meter usage through quotas tied to a user's subscription. Every request sent to a model draws down that allowance, and heavier, more capable models usually consume more of it per task. For developers who lean on AI agents throughout the working day, the rate at which quota disappears is often the practical limit on how much work the tool can do. A feature that stretches the same allowance further therefore matters as much as raw model quality.
Antigravity now offers several Gemini 4 models suited to different engineering jobs. Orchestrator is Google's attempt to make those models work together instead of asking developers to pick one and accept its trade-offs for the entire task.
3. How Orchestrator divides the work
Each model in the pairing has a defined role. Gemini 4 Carbon is aimed at high-level planning and at reviewing output. In an Orchestrator session, it works out how a task should be approached and then checks what has been produced. Gemini 4 Hydrogen is tuned for speed and execution, so it carries out the code edits and runs the tests that verify whether changes work.
The handoffs happen without developer intervention. The system decides when a piece of work can be passed down to Hydrogen and when it should be brought back to Carbon for review. From the user's point of view, the session behaves like a single assistant, even though two different models are contributing at different moments.
This arrangement resembles the way many human engineering teams operate. A senior engineer may design the approach and review pull requests, while implementation is spread across others who can move faster on well-defined work. The value comes from matching the type of task to the resource best suited to it, rather than routing everything through the most expensive option available.
For developers, the main practical change is that they no longer need to switch models manually in the middle of a task to save quota or time. Choosing Orchestrator from the model menu delegates that decision to the platform.
4. What the benchmark numbers show
Google's early performance figures come from Terminal-Bench 4.1, a benchmark that measures how well AI agents complete tasks in a command-line environment. On that test, Orchestrator achieved success rates close to those of Gemini 4 Carbon working on its own.
The efficiency gains are the headline. Compared with Carbon alone, Orchestrator used about 24% less quota on average and completed tasks about 19% faster. In other words, Google is claiming that splitting the work this way gives up little in quality while saving a meaningful share of both resources and time.
These numbers are Google's own and come from a single benchmark. Results on real-world codebases can vary depending on the size of a project, the kind of changes required and how much review a task demands. A job that needs repeated planning and correction may draw more heavily on Carbon, narrowing the savings, while a task dominated by straightforward edits and test runs could benefit more. The figures are best read as an indication of direction rather than a guarantee for every workload.
Even so, the combination of near-equal success and lower quota use is the kind of result that matters to subscribers. If a developer can complete roughly the same amount of work while consuming about a quarter less of their allowance, the effective capacity of their plan rises without any change in what they pay.
5. Who can use Orchestrator, Carbon and Hydrogen
Orchestrator is limited to Google AI Ultra subscribers, Google's top consumer AI tier. Access to Gemini 4 Carbon follows the same line, so the feature that depends on Carbon is available only to those who already have the planning model.
Subscribers on the Google AI Pro plan do not get Carbon or Orchestrator. Their Antigravity access includes the Gemini 4 Argon and Gemini 4 Hydrogen models. Hydrogen is being rolled out gradually to both Pro and Ultra subscribers, so not every account will see it at once.
For Pro users, Hydrogen offers a preview of the execution half of the Orchestrator pairing. They can use the fast model directly, but they cannot yet combine it with Carbon's planning and review.
6. A busy week for Gemini 4
Orchestrator follows closely on the launch of Gemini 4 Carbon earlier the same week. That release also included a preview of /swarm, a multi-agent feature designed for large engineering tasks in which teams of agents work in parallel.
Taken together, the releases show Google building out a layered toolkit inside Antigravity. Carbon provides depth for planning and review, Hydrogen provides speed for execution, Orchestrator coordinates the two inside a single session, and /swarm points toward many agents tackling different parts of a large project simultaneously.
The rapid sequence also suggests that Google sees coding agents as a competitive priority. Development tools are one of the areas where AI models are most directly measured on practical output, and where differences in speed, reliability and cost show up quickly in daily use.
7. Why model routing matters for AI coding tools
Orchestrator is an example of a broader approach often described as model routing, in which each part of a job is sent to the model best suited to handle it. Instead of treating one model as the answer to every request, a routing system weighs the difficulty of a subtask against the cost and speed of the available models.
The logic is economic as much as technical. The most capable models are the most expensive to run, and much of the work in a coding session does not require their full capability. Running tests, applying routine edits and reformatting files are tasks a faster model can usually manage. Reserving the stronger model for the decisions where it makes a difference can preserve quality while lowering cost.
Google's version builds that routing directly into a single session, so the developer does not have to orchestrate it themselves. For subscribers, the practical result is measured in cost per task. A tool that completes the same work with less quota effectively delivers more value for the same subscription fee, which can matter as much as incremental improvements in benchmark scores.
The approach also has implications for how AI providers think about their product lines. Rather than competing solely on a single flagship model, companies can compete on how well their models work together, and on how intelligently their platforms decide which one to use at each step.
8. What Orchestrator changes for subscribers
For Ultra subscribers, the clearest benefit is capacity. Quota limits determine how many agent tasks a developer can run before hitting a ceiling, so a mode that finishes comparable work while drawing about 24% less from that allowance gives heavy users more room before they are throttled. The roughly 19% reduction in completion time adds a second gain, since developers spend less time waiting for an agent to finish before reviewing its output or moving to the next job.
There is also a simplification in day-to-day use. Without Orchestrator, a developer seeking the same balance would need to judge, task by task, whether a step warranted Carbon's depth or could be left to Hydrogen, and then switch manually. That kind of micromanagement is easy to get wrong and adds friction to a tool meant to reduce it. Folding the decision into the platform removes one more choice from the developer's workflow.
For Pro subscribers, the launch draws a sharper line between tiers. They gain access to Hydrogen's speed as it rolls out, but the planning model and the coordination layer that ties the two together remain part of the Ultra package. How much that gap matters will depend on how often a developer's work calls for deep planning and review rather than fast execution alone.
9. What to watch next
The next developments to follow are the continued rollout of Gemini 4 Hydrogen to Pro and Ultra accounts and the progress of /swarm, which is in preview and available first to Ultra subscribers. Whether Orchestrator stays exclusive to Ultra will determine how widely the feature spreads across the platform's user base.
Independent testing will be another area of interest. Google's figures of about 24% lower quota use and about 19% faster completion come from its own Terminal-Bench 4.1 runs. How closely real-world projects match those results will shape whether developers adopt Orchestrator as their default way of working inside Google Antigravity.