Antigravity Thinking Levels: How Reasoning Depth Affects Quota

Editorial illustration of reasoning-depth controls in Antigravity: a thinking-level gauge inside an agent workspace with quota and warning elements. — Reasoning depth is a budget dial, not just a quality dial. Generated hero; not official artwork.

What Thinking Levels Are

Antigravity exposes reasoning depth as a per-run control: Low, Medium, and High. The selector lives in the run configuration and in the /model command surface. The idea is simple — not every task needs the model to think hard, and paying reasoning tokens for mechanical work is waste. The launch posts describe the trade as reasoning depth versus latency; the quota math underneath is the part that matters more.

What the levels are not: a model selector. Thinking levels sit on top of the model choice. You pick Gemini 3.8 Flash (or 3.7 Flash, which remains available), and then decide how hard that model thinks per task. The two controls are independent, and the fast model cadence means you should expect to revisit both.

The Actual Trade

The honest description: higher thinking levels produce more reasoning tokens per turn. Reasoning tokens are output tokens. Output tokens are the expensive side of Flash pricing ($3.75 per 1M at 3.8 launch pricing) and the metered side of Pro/Ultra quotas. So the thinking-level dial is a budget dial wearing a quality costume.

Level	What it does	Best for	When to avoid
Low	Minimal reasoning, fastest responses, lowest output burn	Mechanical edits, renames, doc updates	Latency-sensitive interactive work
Medium	Balanced depth; the implicit middle setting	Feature work, bug fixes, tests	Everyday interactive coding
High	Deepest reasoning, slowest responses, highest output burn	Architecture, hard concurrency bugs, multi-file refactors	One-shot correctness over cost

The /boost announcement is the clearest official statement of this trade. It describes /boost as for “particularly complex tasks and are willing to spend some more” — spend what? Reasoning budget. The wording is casual; the economics are not.

Introducing /boost in Google Antigravity. The Antigravity harness can tackle a wide range of tasks right out of the box, but /boost helps if you are tackling particularly complex tasks and are willing to spend some more.
— @antigravity August 31, 2026

What Quota Actually Measures

Antigravity quotas are tier-based (Pro and Ultra) and metered per model. The launch posts do not publish a per-thinking-level token table, and that absence is the honest caveat in this section. What we can state from official sources: quota resets happen (documented during the 3.8 launch week), resets are per model, and reasoning-heavy runs consume more of the meter than mechanical ones.

The practical implication for planning: if you run a deep-reasoning agent across a full workday, your quota consumption is dominated by output tokens, not input tokens. The cheapest optimization is not switching models — it is matching thinking level to task class, and that is a per-task decision, not a global default.

What We Know vs What We Inferred

Known from official sources: the three levels exist; 3.7 Flash remains available for efficiency-first work; /boost exists for complex tasks at higher spend; quota resets happened during launch week; quota is per model.

Inferred, not documented: the exact token cost per level; the exact quota draw of High versus Medium; whether thinking tokens meter the same as visible output tokens on the plan quota. These are the numbers that would make this post a reference instead of a guide, and they do not exist in official sources at publication time.

The measurement that closes the gap is simple and boring: run one real task at each level against your own workload, record the token counts, and compare. That is the benchmark that actually matters, because it is your code, your repo, and your quota.

Setting Defaults That Survive a Workweek

The failure mode people hit with fast model cadence and thinking levels is a default that silently mismatches their actual work. The defaults that survive are boring:

# Interactive daily driver: Medium on the newest Flash
/model gemini-3.8-flash --thinking medium

# Hard problem: bump depth for this run only
/model gemini-3.8-flash --thinking high

# Mechanical batch: drop depth, keep the model
/model gemini-3.8-flash --thinking low

Three rules make this survive contact with a real workweek. First, set the default at Medium and bump per run, never globally — a global High default silently doubles your quota draw on mechanical work. Second, script your automation with explicit thinking levels, because CI runs do not inherit your interactive defaults. Third, re-verify after every model release — the third Flash in six weeks means your defaults have a shelf life measured in weeks, not quarters.

How /boost Changes This

/boost is the escalation path above the thinking-level dial: the same harness, delegated to focused workstreams with independent verification across iterative rounds. It is for tasks where you would set High and still want more — and it spends accordingly. The launch post describes it as for complex tasks where you are willing to spend more; the practical read is that /boost is a per-task decision, not a default, for the same reason global High is a mistake.

Caveats

No official per-level token table exists. Everything quantitative in this post about quota draw is inference from pricing structure and launch behavior, not documented numbers. The per-level costs are the missing reference, and running your own benchmark is the only way to close it.

Launch-week behavior is volatile. Quota resets during the 3.8 launch window are documented; the reset cadence going forward is not. Do not build automation that assumes this week’s quota behavior persists.

Vendor benchmarks are launch-day claims. Run your own evals before reassigning defaults — especially with a model cadence this fast.

The codebase-analysis demo is the reasoning-heavy workload example worth watching alongside this guide — it shows an agent doing exploration, documentation, and artifact generation in one pass, which is exactly the class of task where thinking levels most affect both quality and quota:

Official Google Antigravity demo: the reasoning-heavy workload class where thinking levels matter most.

Use It If

Use High deliberately: architecture decisions, hard concurrency bugs, multi-file refactors where one-shot correctness beats iteration cost. It is the right tool for exactly those and the wrong default for everything else.

Default to Medium: feature work, bug fixes, and tests hold quality at Medium per the official positioning, and your quota lasts measurably longer across a workweek.

Use /boost sparingly: for tasks where High is not enough and the problem is genuinely multi-workstream. It is the escalation path, not the daily driver — and it draws accordingly.

Measure your own split: one real task at each level against your workload, token counts recorded. That benchmark is the only number that is actually about your quota.

FAQ

What are thinking levels in Antigravity?

A per-run reasoning-depth control (Low, Medium, High) that sits on top of the model choice. Higher levels produce more reasoning tokens per turn — deeper reasoning at the cost of latency and quota.

How much quota does High thinking use?

Google does not publish a per-level token table. Higher levels produce more reasoning tokens, which are output tokens — the metered side. The practical answer is to run one real task at each level against your own workload and compare.

What is the default thinking level?

The launch posts do not document a specific default. Practically: set your interactive default to Medium and bump per run, so a global High default does not silently double quota draw on mechanical work.

When should I use /boost instead of High thinking?

/boost is for complex tasks where you want multi-agent reasoning with independent verification — a step beyond a single High-thinking run. It spends more, so treat it as a per-task escalation rather than a default.

Do thinking levels apply to 3.7 Flash too?

Yes. Thinking levels are a control in Antigravity, independent of which model you select. 3.7 Flash remains available specifically for efficiency-first workflows, where lower thinking levels compound the savings.

Does the thinking level persist across agent runs?

Selection is per run. Scripted and CI runs need the thinking level set explicitly in run configuration, because they do not reliably inherit your interactive defaults.

Sources

Official sources first. All accessed September 2–9, 2026.