Skip to content

Google Gemini

export GEMINI_API_KEY=...

from promptcanary import LiteLLMProvider

provider = LiteLLMProvider("gemini/gemini-3.5-flash", temperature=0.0)
Tier Model string Notes
Flagship gemini/gemini-3.1-pro Highest capability, strongest reasoning.
Balanced gemini/gemini-3.5-flash Good default — fast and capable.
Fast / cheap gemini/gemini-3.1-flash-lite Cheapest Gemini 3 tier, good for frequent checks.

CLI Usage

promptcanary run --provider gemini/gemini-3.5-flash --save-baseline
promptcanary compare --provider gemini/gemini-3.5-flash --fail-on-drift

Multi-Provider Example

A common pattern is running the same suite across OpenAI, Gemini, and a free local model to spot provider-specific drift versus universal harness issues:

from promptcanary import CanarySuite, LiteLLMProvider

suite = CanarySuite.from_yaml("canary.yaml")

for model_id in [
    "openai/gpt-5.4",
    "gemini/gemini-3.5-flash",
    "ollama/qwen3.6:27b",
]:
    provider = LiteLLMProvider(model_id, temperature=0.0)
    result = suite.run(provider, show_progress=False)
    print(f"{model_id:35s}  score={result.overall_score:.1%}")

If only one provider regresses, the issue is provider-specific drift. If all three regress simultaneously, suspect your suite's prompts or your own harness instead.