Google Gemini

Gemini's headline act is the context window: gemini-3-pro takes one million tokens of input — whole-repo comprehension, long-log analysis, giant-diff review — at a mid-tier price. Stella speaks the native Gemini generateContent dialect against https://generativelanguage.googleapis.com/v1beta.

Setup

export GEMINI_API_KEY="..."

GOOGLE_API_KEY works as an alias — Stella checks both, GEMINI_API_KEY first. Gemini sits sixth in auto-detection. Prefer a file? Store the key in ~/.config/stella/credentials.toml, or rename the env var Stella reads with providers.gemini.api_key_env — see overriding the key source.

Selecting models

The provider id is gemini; the default model is gemini-3-pro.

stella --model gemini/gemini-3-pro run "summarize why the nightly build broke"
export STELLA_MODEL=gemini/gemini-3-pro

The same models are also reachable through Vertex AI if your organization runs inference inside Google Cloud — same wire dialect, project-scoped credentials.

Thinking

Stella's effort setting maps to Gemini's thinkingLevel — low or high. One family quirk to know: Gemini 3 cannot fully disable thinking. Setting reasoning: "off" gets you the minimum thinking level, not zero — budget accordingly on latency-sensitive roles.

In the pipeline

gemini-3-pro is a solid worker and a good judge, but its distinctive fit is any role that needs the 1M window — a worker on a task whose context simply does not fit elsewhere, or a judge that must read an entire subsystem to verify a change. See Examples for profiles and agent_engine_config for per-agent routing.

Catalog pricing

List prices from Stella's seed catalog — verify with Google; your local catalog re-syncs from models.dev via stella models refresh. See the Model guide for the full cross-provider table.