How to Reduce Antigravity Token Usage

Part of the reduce AI coding agent token usage pillar. This page is Antigravity-specific: what its quota actually measures, and what you can do about it.

How do you reduce Antigravity token usage?

Give the agent less to read. Antigravity's rate limits are, in Google's own words, "correlated with the amount of work done by the agent, which can differ from prompt to prompt" — so your quota is not spent by asking, it is spent by how much the agent has to chew through to answer. A short prompt that sends it across forty files costs more than a long prompt that points it at one. That single sentence in the documentation is the most useful thing Google has published about Antigravity's economics, and it changes what optimisation means here. You are not rationing prompts. You are rationing agent work, and agent work is mostly reading.

How does the Antigravity quota actually work?

Three tiers, two refresh cadences. Per the official plans documentation:

Google AI Ultra — the highest quota, refreshed every five hours, highest weekly rate limits, plus access to third-party models.
Google AI Pro — a high quota, also refreshed every five hours until the weekly limit is reached, with a higher weekly ceiling.
Everyone else — a "meaningful" quota refreshed weekly, with a weekly rate limit.

Every plan gets the same baseline access: Gemini 3.1 Pro, Gemini 3.5 Flash and other Gemini Enterprise Agent Platform models as the core agent model, unlimited Tab completions, and all product features including Scheduled Tasks and the CLI. Two things follow. First, the five-hour refresh on Pro and Ultra behaves like a rolling window, not a monthly allowance — the same shape as Claude's limit reset, and it rewards spreading heavy work across windows rather than front-loading a day. Second, Google states plainly that these limits "are primarily determined to the degree we have capacity" and are "subject to modification". Plan around the mechanism, not around a number you measured last month.

What happens when you run out?

On Pro and Ultra, purchased AI credits cover overage — if you let them. Credits are consumed at standard Gemini Enterprise Agent Platform consumption pricing, and their use is governed by an "AI Credit Overages" user setting. That setting is worth finding before you need it. It is the difference between an agent that stops when the quota is exhausted and one that quietly keeps going against your credit balance. Neither behaviour is wrong, but only one of them is a surprise.

How do you see your Antigravity usage?

/usage, or its alias /quota, in the CLI prompt box. It refreshes your model configuration and quota status from the backend and opens an interactive TUI panel showing a breakdown of your usage limits and remaining requests or tokens for each supported model — Gemini 3.5 Flash and Gemini 3.1 Pro among them. The panel triggers a fresh check against both disk and the backend service when opened, so the numbers are current rather than cached. Navigation is arrow keys or J/K, PgUp/PgDn by page, G and Shift+G for top and bottom, Esc or Q to close. Read the per-model split first. If one model is draining far faster than the others, that is a routing question — which work goes to Pro versus Flash — before it is an optimisation question. Be clear about what /usage cannot tell you. It is a quota meter: it reports what remains, not what consumed it. It will not tell you that eleven of your last twenty tool calls re-read the same four files. For that you have to watch a live session, or put something between the agent and its tools.

Why does an agent-first IDE burn quota faster?

Because the agent decides what to read, and it errs toward reading more. Antigravity is built around an agent manager rather than an editor with completions bolted on, which is the point of the product — and it means more of your consumption is decided by the agent's judgement rather than by your keystrokes. The patterns are the same ones that inflate every agentic tool, and none of them appear as a line item:

Whole-file reads. A 2 000-line file opened to change four lines bills the other 1 996.
Command output. Test runs, install logs and diffs get piped back in full, mostly noise.
Tool manifests. Every connected MCP server ships its tool definitions before your first message, used or not.
Re-reads. The same file entering context four times in one session costs four times, and it is the failure mode you are least likely to spot.

Google's own framing confirms why this matters more here than on a per-message plan: "you may get many more prompts if your tasks are more straightforward and the agent can complete the work quickly, and the opposite is also true."

What is free, and what is metered?

Tab completions are unlimited on every plan. The agent is not. That is stated in the baseline quota for all tiers, and it is the most under-used fact about Antigravity's economics. It draws a line most people never think about. Completions are cheap to serve, so Google gives them away; agent runs are expensive, so they are rationed. Every task you complete by typing with Tab assistance instead of delegating to the agent costs you nothing against the quota that actually binds. This is not an argument for doing everything by hand — the agent exists because it is faster on work that spans files. It is an argument for noticing which of the two you are reaching for. A rename you could Tab through in fifteen seconds, handed to the agent, spends metered quota on something the free path does perfectly well. Over a week that adds up to a meaningful share of a five-hour window. The same logic applies in reverse to Scheduled Tasks, also included on every plan: work that genuinely needs the agent is better batched into a well-specified autonomous run than dripped out interactively, because each conversational turn replays context that a single run pays for once.

Does the plan you are on change the terms?

Individual and team access run on different agreements, and it matters for more than billing. Google states that Antigravity is available to individual accounts under terms derived from Google's terms of service, and to teams under Google Cloud terms through the Gemini Enterprise Agent Platform. Rate limits and model availability both differ according to your Google AI plan rather than being a property of Antigravity itself. If you are evaluating it for a team on the basis of what it does on your personal account, confirm which models your organisation's plan actually exposes before building a workflow on one of them.

What are the biggest levers?

Scope the task before you send it. Naming the files or the directory turns an open-ended search into a bounded one. This is the highest-leverage habit on any agent-first tool and it costs nothing. Route by model deliberately. Flash and Pro do not consume the same quota. Reserve the heavier model for work that needs it, and check the split in /usage rather than assuming. Prefer autonomous, well-specified runs over interactive back-and-forth. Each conversational turn replays context. A single well-scoped task that the agent completes in one pass moves less text than six rounds of clarification. Disconnect MCP servers you are not using. Their manifests are a standing cost on every turn, not a cost at call time.

How to apply this today

Run /usage before changing anything. Ten minutes with your own per-model numbers beats every generic recommendation, including the ones on this page, because the distribution of waste differs by codebase.
Find the AI Credit Overages setting and decide deliberately whether you want the agent to spend credits past the quota.
Scope your next five tasks explicitly — name files or directories — and compare the quota drain against your usual pattern.
Audit what is connected. Every MCP server you are not actively using is paying rent on every turn.

If you would rather not do steps 3 and 4 by hand on every session, that is the gap Tokenade fills. It sits between the agent and its tools, replacing whole-file reads with structure-first reads and semantic search, compressing command output, and deferring MCP manifests until they are actually used. Antigravity is one of the agents it hooks natively, and the free tier covers a lot of solo work outright.

What goes wrong (anti-patterns)

Treating the quota as a prompt count. It is a work meter. Optimising the number of messages you send while leaving the agent free to read the repository on each one changes almost nothing. Front-loading a day on a five-hour refresh. On Pro and Ultra the window rolls. Burning the whole allowance before lunch does not buy you the afternoon. Leaving credit overages on by default without knowing it. The quota stops being a ceiling the moment credits cover it, and the first time you notice is usually the balance. Capping output. Output is the expensive side per token, but agent traffic is overwhelmingly inbound. It is the lever people reach for first and the one that pays least. Assuming last month's limits. Google states the limits are capacity-dependent and subject to modification. Re-check /usage rather than trusting a number you remember.

See also: