TabbitBlog

Google Antigravity Orchestrator: how model routing works

Learn how Antigravity Orchestrator can hand work between Gemini models, who can access it, and how to assess its quota and speed claims.

In this article
  1. Key takeaways
  2. Google Antigravity Orchestrator at a glance
  3. How a model handoff can work
  4. When should you select Orchestrator?
  5. Access and quota are separate from Gemini API use
  6. What the 24% and 19% figures can tell you
  7. How to evaluate it fairly
  8. Orchestrator versus manual model routing
  9. When the workflow crosses into the browser
  10. Final verdict

Google Antigravity Orchestrator is a model option for delegating different stages of a coding task to different Gemini models. Google's launch example uses Gemini 4 Carbon to inspect a codebase and plan, Gemini 4 Hydrogen to make changes and run tests, then Carbon again if the work hits friction. It is meant to manage those handoffs inside one session.

The short verdict: Orchestrator is worth testing when a task needs both careful planning and quick execution. Google reports lower quota use and faster completion than Carbon alone on a Terminal-Bench test. That is a benchmark result, not a promise for your repository. One Antigravity user framed the practical uncertainty well: “the less expensive agents are just simply not as accurate.” The question is whether Orchestrator can keep quality while reducing review and rework on your task. Read the discussion.

Key takeaways

  • Orchestrator is dynamic model routing within one task. Google says it can hand work between Carbon and Hydrogen as the task changes.

  • The handoff example is a useful mental model, not a public specification of every trigger. Google has not described all controls or the exact switching logic in the launch post.

  • Google reports 24% less quota use and 19% faster completion than Carbon on Terminal-Bench. Its product post labels the result Terminal-Bench 4.1, while its Gemini announcement labels it 4.0. That discrepancy is unresolved.

  • A quota percentage is not the same as API token cost. Antigravity account usage and Gemini API billing are different meters.

  • Access is plan-specific and can change during rollout. The announcement initially described Orchestrator for AI Ultra and Gemini for business. Check your own model selector and current plan details.

  • Routing helps only if the time and quota saved exceed the cost of handoffs, review, and corrections.

Google Antigravity Orchestrator at a glance

QuestionWhat Google saysWhat remains uncertain
What is it?A model dropdown option that pairs Hydrogen with stronger models for long-horizon workExact routing triggers for every task
Which models?Example trajectory switches between Gemini 4 Carbon and HydrogenNo promise that every step follows the same sequence
How does the handoff work?Carbon can explore and plan, Hydrogen can edit and test, Carbon can resume if Hydrogen gets stuckUser controls for forcing or locking each step are not fully described
What is the goal?Balance quality, efficiency, and latency in a single sessionOutcome varies by task and repository
Quota claimGoogle reports 24% lower average quota use than Carbon on Terminal-BenchNot a guaranteed plan saving or conversion to API tokens
Speed claimGoogle reports 19% faster average completionNot a guarantee of shorter end-to-end work after review
Who can access it?Launch notice says AI Ultra and Gemini for business; rollout could changeEligibility by region/account and later plan additions must be checked live

How a model handoff can work

Google Antigravity announcement page for Gemini 4 Hydrogen and Orchestrator.
Google Antigravity announcement page for Gemini 4 Hydrogen and Orchestrator.

A static router looks at a prompt once and picks a model for the whole job. Orchestrator is described differently. Google says the model choice can change as a task develops, because the model that can plan well may not be the most efficient model for every edit or test.

The public example has three stages:

  1. Explore and plan. Carbon reads the relevant parts of a repository, resolves ambiguities, and creates a plan with conditions for replanning.

  2. Implement and verify. Hydrogen handles focused edits, package updates, and tests.

  3. Recover when needed. If Hydrogen hits a blocker, control can return to Carbon to inspect the new state, revise the plan, and continue.

That sequence explains the idea, but it should not be read as a deterministic contract. The Antigravity product post describes an example path. It does not publish every routing trigger, a guarantee that every run uses both models, or a complete set of user controls for locking each handoff. Look at current documentation and the task trace exposed by your version before building a process around a presumed rule.

There is another Google-published detail worth keeping in view. The launch post says developers can build an explicit multi-model pipeline through the Antigravity SDK when they want programmatic control over which model handles each role. That is a different approach from selecting Orchestrator in the UI. If you need deterministic stages for an automated system, compare the SDK's documented behavior and terms rather than assuming the dropdown gives the same control.

When should you select Orchestrator?

The strongest candidate is a task with changing work: broad exploration, a concrete plan, repeated edits, tests, and occasional replanning. The value may be lower for a short task with one obvious answer. A single model or a manual sequence can be simpler when you already know which model should handle each stage.

Workflow signalOrchestrator may help when…Measure thisWatch for
Large codebase discoveryThe task needs broad read-only exploration before editsRelevant files found, missed dependencies, planning timeToo much planning before a small change
Routine implementationMany concrete edits follow a stable specPassing tests, elapsed time, correction countFast edits that create review work
Unclear requirementsNew information may require replanningReplans that improve the outcomeRepeated handoffs with no clearer plan
Test failure recoveryA lower-cost execution model may get stuck and need stronger diagnosisRecovery success and total quotaA handoff that repeats the same failed approach
Short, isolated changeWork fits one model and one short tool sequenceCompletion and total timeRouting overhead exceeds the task itself
Strict reproducibilityYou need to know which model performed each stepTrace visibility and repeatabilityHidden transitions that make results hard to audit

A commenter in the same subreddit described the review problem: a cheaper executor may create enough errors that a stronger model spends the savings fixing them. A separate rollout thread raised a basic eligibility question: does Orchestrator work on a student offer? That is a reason to check your own account, not a source for plan policy. These questions point to the right measurement: finished work after review, not the first model call. Read the task-quality discussion and the rollout question.

Access and quota are separate from Gemini API use

The initial product rollout notice says Gemini 4 Hydrogen is rolling out to paid Google AI Pro and Ultra subscribers in Antigravity, while Orchestrator is rolling out to Google AI Ultra subscribers. It also lists access for Gemini for business. This differs from the Hydrogen API and AI Studio availability listed in the broader Gemini announcement and model card. The channels have separate eligibility and usage rules.

Launch windows move quickly. A Reddit thread claimed that a named Orchestrator model had appeared on AI Pro, while users in the same thread asked whether student offers counted. That forum report is not official plan policy. Check the current model picker and Antigravity's documentation for your account before treating an early rollout observation as a general entitlement.

Also separate these two cost questions:

  • Antigravity quota: the product's plan usage meter. Google’s 24% claim compares quota use on its benchmark. It does not say that your account will always spend 24% less.

  • Gemini API bill: model input and output token rates billed for API requests. If you use the API, calculate cost from the current rate card and your own token counts. It does not translate directly from Antigravity quota.

The difference matters because long-running agent tools carry context and perform multiple calls. A lower-priced model on one API request does not prove that an orchestrated task will use less of a subscription quota. Likewise, faster completion does not mean lower API billing if the run uses more tokens or retries.

What the 24% and 19% figures can tell you

Google's Antigravity post reports near-Carbon task success with 24% less quota use and 19% faster completion on Terminal-Bench 4.1. The Gemini 4 announcement gives the same average percentages but labels the benchmark Terminal-Bench 4.0. The official pages disagree on the test version, and this article cannot resolve that difference. Treat the reported result as Google's benchmark claim and verify the methodology before citing it in a formal evaluation.

The figures are a reason to test Orchestrator, not a forecast for every team. A benchmark controls its task set and environment. Your result depends on repository size, test quality, network access, context, plan, concurrent load, model rollout, and the amount of human review. A 19% faster agent run that needs a long manual correction is not a 19% faster task.

If you compare it with a single-model run, measure these separately:

end-to-end time = agent time + tool wait time + human review time + correction time
successful-work efficiency = accepted task completions ÷ quota consumed

These simple measures make the decision clearer than a model label. Count a task as complete only when the same tests, acceptance criteria, or source checks pass for both approaches.

How to evaluate it fairly

You do not need a large benchmark suite to learn whether routing helps your own team. Start with several tasks that resemble your real work and use the same conditions for Orchestrator and the best single-model option available in your account.

  1. Pick 8–12 tasks with a clear acceptance check. Include a small task and a longer task that requires discovery, edits, and verification.

  2. Give both approaches the same repository state, prompt, tool permissions, network access, and test commands.

  3. Randomize which approach runs first, then repeat each task several times if quota allows. Do not compare one run from a quiet hour with another from a busy period.

  4. Record the task success criteria, total wall time, quota before and after, number of handoffs or retries visible in the trace, review minutes, and corrections.

  5. Count failed runs and regressions. A fast first pass that fails tests is not a completed task.

  6. Compare the accepted output and total human effort. If Orchestrator hides the model transition details, record that as a limitation in the evaluation.

Use a small decision rule: keep Orchestrator for a task class only if it produces a useful gain in completion, time, or quota without pushing review work above the value of that gain. If results are mixed, keep it as a choice for specific tasks rather than making it the default.

This kind of test is relevant to any agentic browser or coding workflow, but the exact model availability and meter belong to Antigravity. A browser agent also adds page state and login context to the test; see the guide to browser automation and the overview of browser agents before comparing models across different environments.

Orchestrator versus manual model routing

Orchestrator saves the user from deciding every model handoff, which can make it easier to run a task from one prompt. Manual routing gives a person or script more direct control over the sequence. Neither is automatically more efficient.

Use Orchestrator when task stages are hard to predict and the convenience of automatic switching is valuable. Use manual routing when the roles are stable, the transitions need to be auditable, or you want a particular model to review each change. Use one model when a task is small enough that coordination and extra context would be waste.

The label “multi-model” does not mean “multiple independent agents” by itself. Google's example is a single trajectory that can return to Carbon. Its separate /swarm description is about parallel agents and worktrees. Do not use a swarm benchmark to predict Orchestrator, or assume a parallel-agent quota rule applies to a sequential model handoff.

When the workflow crosses into the browser

A code task that reads web sources adds another variable: the browser and its page context. Keep those fixed in the comparison. The AI browser guide explains the browser environment, browser-use covers browser-driven tasks, and the guides to research browsers and deep research describe adjacent web-research workflows. These are useful context, not evidence that Orchestrator routes browser tasks.

Final verdict

Google Antigravity Orchestrator is a managed way to move between Gemini 4 Carbon and Hydrogen as one coding task changes from planning to execution and recovery. Its reported quota and speed gains make it worth testing on long tasks that genuinely need both model strengths.

Use the feature when you can verify access and measure its complete result. Keep review quality, rework, and the plan's actual quota meter in view. If model choice needs to be predictable, or the task is short, a single model or an explicit pipeline may be easier to reason about. The Tabbit vs Antigravity comparison covers the broader product distinction; this article is about Antigravity's Orchestrator option.

FAQ

What is Google Antigravity Orchestrator?

Orchestrator is a model option in Google Antigravity that can hand parts of one task between Gemini 4 Carbon and Gemini 4 Hydrogen. Google describes it as dynamic routing across a task, rather than choosing one model for the entire run.

Which models does Antigravity Orchestrator use?

Google's November 2026 launch post describes Orchestrator working between Gemini 4 Carbon and Gemini 4 Hydrogen. The post does not promise a fixed sequence for every task or list all future model combinations.

Who can use Google Antigravity Orchestrator?

Google's launch announcement initially described Orchestrator as rolling out to Google AI Ultra subscribers and Gemini for business. Rollouts can change, so check the model picker and current plan details in your account.

Does Orchestrator reduce Antigravity quota use?

Google reports 24% lower average quota use than Carbon on a Terminal-Bench evaluation. That result is specific to Google's benchmark and does not guarantee the same saving on your project.

Is Google Antigravity Orchestrator faster?

Google reports a 19% faster average completion time on its Terminal-Bench result. Your task, network, plan limits, and any review or rework can change the total time.

Can I choose or inspect every Orchestrator model handoff?

Google explains example handoffs but does not fully document every trigger or user control in the launch post. Check current Antigravity documentation and inspect the task trace available in your version before relying on a specific routing rule.

Take the next step

Let Tabbit work alongside you.

Research across tabs, automate repetitive browser work, and keep every piece of context within reach.