Checked 30 August 2026
Every figure on this site is taken from Google's own pages and carries the date it was read. Nothing here is a benchmark we ran.
Google Antigravity, written down and dated.
An independent reference to the agent-first development platform Google launched in November 2025 — what the surfaces actually are, which models you can pick, what the quotas cost, and where the published numbers stop.
Gemini 3.7 Flash
Select a model to read its published profile.
- Provider
- —
- Quota pool
- —
- In picker
- —
- Per 1M in / out
- —
Bars render Google's own three-step words — Low, Medium, High, Fast — from the models table, not measurements taken by anyone here. A blank bar means Google publishes no rating for that model. Source: antigravity.google/docs/models, read 30 August 2026.
Directory
Four rails, grouped the way Google's own documentation groups them. Names are Google's; the one-line descriptions are ours.
Source: antigravity.google/docs and antigravity.google, read 30 August 2026.
A — first rail
Antigravity is not one program, and that is the part most write-ups get wrong.
At launch there were two views: an editor and a manager. By August 2026 the product page lists five separate surfaces onto the same agents. Which one you use changes almost nothing about what the agent can do and almost everything about how you watch it work.
Surfaces — Five ways in, not one application
5 entriesAntigravity IDE
The full agentic editor: tab autocompletion, natural-language commands and an agent that operates across the editor, terminal and browser.
Antigravity 2.0
Google's phrase for it is a command centre — managing several local agents running in parallel rather than one conversation at a time.
Antigravity CLI
A terminal-first surface for the same agents, for people who would rather not leave the shell.
Antigravity SDK
A Python kit for prototyping custom agents against the platform.
Antigravity Extensions
Extension support, so an existing editor set-up carries over.
Artifacts — What the agent hands back instead of a log
5 entriesImplementation plans
The agent writes the plan before it writes the code, so the thing you review is the intent rather than the diff.
Task lists
Work broken into steps you can follow while it runs.
Screenshots
Visual evidence that the change did what it said, captured by the agent.
Browser recordings
A replay of the agent driving the browser — the strongest form of verification the platform produces.
Artifact review
Comments and feedback can be left on any artifact to steer the next pass.
Customisation — What you can bolt on
6 entriesMCP servers
Model Context Protocol connections to tools and data outside the editor.
Skills
Packaged procedures the agent can call for a particular kind of task.
Rules
Standing instructions that apply to a project without being retyped.
Plugins
Bundles that install a set of the above together.
Hooks
Points where your own code runs in response to what the agent does.
Sidecars
Companion processes running alongside the agent.
Agent capabilities — How far it is allowed to go
6 entriesSlash commands
Named shortcuts for specialised tasks, typed into the same chat input.
Teamwork
Several agents working the same repository without treading on each other.
Permissions
The controls that decide what the agent may do unattended.
Subagents
An agent delegating part of a task to another agent.
Worktree mode
Starting a run in a fresh git worktree instead of your working copy — the difference between an experiment and an edit.
Remote control
Reaching a running agent from somewhere other than the machine it is on.
Index — models
Gemini models and third-party models draw down different allowances. The interface shows each as a five-hour figure and a weekly one.
Source: antigravity.google/docs/models, read 30 August 2026.
B — the picker
Seven reasoning models, two quota pools, one image tool.
The interesting detail is the pooling. Gemini usage and third-party usage are metered separately, so exhausting Claude does not stop you working — it stops you working that way. Google's plans page puts access to third-party models on the Ultra tier.
Google publishes speed, latency and intelligence words for its own models and not for the third-party ones. We have left those cells blank rather than filling them in.
Prices shown are the published per-million input / output token figures where Google states them. Gemini 3.7 Flash is on an introductory price through the end of 2026, rising to $1.50 / $7.50 on 1 January 2027. Source: Gemini 3.7 Flash in Google Antigravity, 13 Aug 2026.
C — the sequence
What a run actually looks like
Six steps, in the order Google's getting-started page puts them, with the thing each step is easy to get wrong noted underneath.
Source: antigravity.google/docs, read 30 August 2026.
The file comes from Google's domain
Builds for macOS, Windows and Linux live on antigravity.google. If the setup offers Replace, that is the expected prompt.
This site links to that page and mirrors nothing. Any other site handing you an Antigravity build is not Google.
Open a project
The folder icon with a plus, in the left sidebar. A project is a local folder or a git repository.
Nothing is uploaded by opening a project; the agent works against the files on your machine.
Say what you want, not which tool to run
The goal goes in the chat input in plain language. This is the whole premise: you are handing over a task, not a command.
Google's own framing is operating “at a higher, task-oriented level”.
Choose local or a new worktree
Local mode works in your checkout. New worktree mode branches into a fresh git worktree.
Worktree mode is the one to reach for when you are not yet sure the agent should be trusted with the change.
Read the artifacts, not the log
The agent produces a plan, a task list, screenshots and browser recordings. Comments left on any artifact steer the next pass.
This is the part that distinguishes it from autocomplete: the unit of review is the deliverable, not the token stream.
Watch the two limits
The interface shows a five-hour figure and a weekly one, tracked separately for Gemini and for Claude and GPT.
Google says limits are “correlated with the amount of work done by the agent, which can differ from prompt to prompt” — so they are not predictable from a prompt count.
Before / after
Google's published comparison of Gemini 3.7 Flash against 3.6 Flash, 13 August 2026. Upper bar is 3.6, lower bar is 3.7.
Each bar is drawn on that benchmark's own scale, so the three rows are not comparable with each other.
Source: Gemini 3.7 Flash in Google Antigravity, 13 Aug 2026.
D — the evidence
One model generation of difference, measured by the vendor.
These are Google's numbers for Google's model, which is the correct amount of scepticism to hold about them. They are still the only paired figures published, and they are specific enough to check.
The number worth sitting with is the second one. On FrontierCode both cuts fail more tasks than they finish. An agent platform is not a reason to stop reading the diff.
Google's headline figure for the release: resolving real software engineering issues end to end.
A harder coding set, where both cuts still fail the majority of tasks. Worth keeping in view before delegating anything unattended.
Published for 3.7 Flash only, so there is no paired 3.6 figure to show. An Elo rating is relative to the field, not a percentage.