Skip to content
Docs / Platform / Antigravity helpers

Antigravity helpers

What the Antigravity subscription can do for any MaxiMind conversation, tested on 03–04/10/2026. Use it before paid keys and before your own Claude account.

Antigravity runs on a subscription: there is no per-call bill. The quota is shared by every conversation and resets on its own schedule, so split big jobs and don't waste it.

agy models                                   # list the models this account has
agy --model <model> -p "<prompt>" < /dev/null   # one prompt, printed answer

Always start a helper prompt with: "Do not message other conversations or people. Change only the files named here."

Models

Model id Use it for
claude-opus-5-5-high / -medium / -low Claude-quality building, reviewing and writing, without using a Claude account (Opus 5.5 answered in 12 s in our test)
claude-sonnet-5-5-high / -medium / -low Faster Claude work (Sonnet 5.5 answered in 5 s)
gemini-4-argon-high / -medium / -low Review, structured JSON from long text, long documents given as a FILE. Replaces Gemini 3.1 Pro since 13/10/2026 (the CLI no longer lists Pro). Not every account carries it; MaxiMind picks one that does. High is the top level: there is no xhigh or max. Not for web research: slow, and it mixed in an outdated number (14/10 test)
gemini-3.8-flash-*, gemini-3.7-flash-*, gemini-3.6-flash-* Quick, cheap-on-quota tasks
gpt-oss-120b-medium Open model, rarely needed

Listening to audio (measured with Gemini 3.1 Pro)

These numbers are from Gemini 3.1 Pro, which the CLI stopped listing on 13/10/2026. Listening has NOT been re-measured with Gemini 4 Argon: until it is, use Whisper (next section). With Gemini 3.1 Pro, Antigravity read an audio file in the working folder and wrote what was said:

  • English test sentence: word for word, 15 s.
  • Vietnamese test sentence: word for word with all diacritics, 24 s.
  • Accent judgement is NOT reliable: it called a standard macOS voice "Southern". Never use it to approve a dialect.
# As it was run in September. The CLI now refuses this model id; measure gemini-4-argon-high before relying on it for listening.
agy --model gemini-3.1-pro-high -p "Listen to clip.m4a in this folder and write exactly what is spoken, with Vietnamese diacritics." < /dev/null

Which tool listens? (Phi's rule, 04/10)

  1. Whisper first: MaxiMind speech-to-text (/jobs, see Speech-to-text). Free, no limit, runs on our Mac.
  2. Antigravity Gemini 3.1 Pro as a second opinion, for example on clips Whisper flags. It's fine to use, but the quota is shared.
  3. Never paid Google keys for listening. Paid Gemini/Google keys may only make text-to-speech audio. Spend is capped at a $20/day warning and a $50/day stop per key.

Why this page exists

On 03/10 a podcast job listened to its own takes with the paid gemini-2.5-pro API, an older, weaker model billed per call (2,474 calls in about 30 hours), while the newer Gemini 3.1 Pro was available on the subscription. Check this page before choosing a model for any job.