Supported Models & Reasoning Controls

The Antigravity CLI (agy) serves the Gemini 3.8 / 3.7 / 3.6 Flash families and Gemini 3.1 Pro, plus three models that are not Gemini at all and sit on their own quota.

This page lists what you can pass as model, and what effort does on top of it. Run gemini-models for the live list from your own install — a written list can go stale, that one cannot.

Which model runs when you name none

There is no default. gemini-ask and gemini-brainstorm send no --model flag at all when you leave model unset, so agy answers on whatever model it is itself configured to use.

Two tools differ. gemini-plan pins gemini-3.8-flash-high when you name none. gemini-image has no model parameter at all.

The models you can name

Pass any of these as model to gemini-ask, gemini-plan or gemini-brainstorm.

Model Identifier	Alias	Best For
gemini-3.8-flash-high	flash	The usual pick. Code review, large refactoring, Q&A, and planning.
gemini-3.8-flash-medium	-	Everyday tasks with balanced reasoning depth.
gemini-3.8-flash-low	-	Summaries, syntax checks, formatting.
gemini-3.7-flash-high	-	Previous generation, high reasoning.
gemini-3.7-flash-medium	-	Previous generation, balanced.
gemini-3.7-flash-low	-	Previous generation, minimal thinking.
gemini-3.6-flash-high	-	Older generation, kept for reproducing earlier results.
gemini-3.6-flash-medium	-	Older generation, balanced.
gemini-3.6-flash-low	-	Older generation, minimal thinking.
gemini-3.1-pro-high	pro	Deep architectural reasoning. Algorithm design, math proofs, security boundary audits.
gemini-3.1-pro-low	-	Pro-grade knowledge retrieval with minimal thinking.

Non-Gemini models

agy also serves three models that are not Gemini, and they are billed against a separate quota bucket. Gemini quota being spent does not stop them, and vice versa.

Model Identifier	Notes
claude-opus-4-6-thinking	Extended-thinking Claude Opus.
claude-sonnet-4-6	Claude Sonnet.
gpt-oss-120b-medium	Open-weights GPT-OSS 120B.

They take the same model parameter as everything else:

{
  "name": "gemini-ask",
  "arguments": {
    "prompt": "@src/backends/agy.ts review the fallback ladder",
    "model": "claude-opus-4-6-thinking",
    "mode": "plan"
  }
}

Reasoning effort

Two things control thinking depth, and they are independent.

One is the -high / -medium / -low suffix baked into the model id. The other is the effort parameter, which the relay forwards to agy as --effort — but only when the installed build advertises that flag, which is the rule for every flag the relay sends. The relay does not reconcile the two: set both and agy decides.

{
  "name": "gemini-ask",
  "arguments": {
    "prompt": "Audit @src/utils/geminiExecutor.ts for path traversal and symlink escapes.",
    "model": "gemini-3.1-pro-high",
    "effort": "high",
    "includeUsage": true
  }
}

• effort: "high": maximum reasoning tokens, for rigorous edge-case analysis, distributed systems design, or security audits.
• effort: "medium": balanced reasoning for refactoring and feature implementation.
• effort: "low": minimal thinking overhead for summaries, classification, and translation.

gemini-ask and gemini-brainstorm send no --effort unless you pass one; gemini-plan defaults to "high". The parameter is documented for the Gemini 3.8 Flash, 3.7 Flash and 3.1 Pro families.

Pass includeUsage: true to get the input, output and thinking token counts back with the answer. It is deliberately ignored when you also set a jsonSchema: the structured body is what you run JSON.parse on, and a trailing 📊 [Tokens: …] line would make it unparseable.

Finding out what you can reach

The gemini-models tool asks the CLI itself (agy --output-format json models) and prints what came back, alongside the active backend's capabilities:

"Run gemini-models to see what models and reasoning options are available."

For quota, gemini-doctor runs agy -p "/usage" --output-format json, which agy's command layer answers without a model turn and therefore costs no tokens. It reports each quota group with the percentage left and its reset time, which is where the split between the Gemini bucket and the Claude/GPT bucket is visible.

For AI agents — aliases, the offline list, and the legacy backend

Aliases. Only two targets exist, and the relay resolves the id before it reaches agy:

• flash, gemini-flash, gemini-3.8-flash → gemini-3.8-flash-high
• pro, gemini-pro, gemini-3.1-pro → gemini-3.1-pro-high

The retired gemini-2.5-flash and gemini-3.5-flash ids resolve to the first, and gemini-2.5-pro to the second, so an old call does not break — but they are no longer models you can reach on this backend.

When the live query fails. The backend runs agy --output-format json models; if that call fails or comes back empty it returns a built-in list of eight ids instead — gemini-3.8-flash-high, gemini-3.8-flash-medium, gemini-3.8-flash-low, gemini-3.7-flash-high, gemini-3.7-flash-medium, gemini-3.7-flash-low, gemini-3.1-pro-high and gemini-3.1-pro-low. A list with no 3.6 Flash and no non-Gemini models in it therefore means the live query did not answer, not that the other models are gone.

The legacy Gemini backend. With GEMINI_MCP_BACKEND=gemini the relay drives the retired Gemini CLI instead. That backend offers only gemini-2.5-pro and gemini-2.5-flash, and supports no reasoning effort, no --json-schema and no execution modes — runWithBackend drops those arguments and says so in a ⚠️ notice rather than silently ignoring them. The automatic Pro-to-Flash retry on a quota error also lives only there; on agy a quota error is surfaced verbatim instead.