Proxies for Gemini and the Google AI API

Why proxy Gemini / Google AI API traffic

1. Regional eval on a provider that routes by geography. Google's inference infrastructure routes APAC traffic through APAC POPs (asia-northeast1, asia-southeast1, etc.) when the client origin is in-region. Measuring whether the response differs from a US-cloud origin vs. an in-region residential is methodology.

2. Vertex AI multi-region workloads. Teams using Vertex for production inference sometimes want to test how their own pipelines behave when the request appears to originate from a different region than the inference endpoint.

3. Google-specific content policy. Google applies country-specific content policy (particularly for image generation and content filtering) that shows up in eval from different origins.

4. Multilingual eval. Gemini handles many languages; pairing each language with its primary origin country (via residentials) gives the authentic eval origin.

What this page isn't

• Not for circumventing Google's restrictions on Gemini availability.
• Not for bypassing API quota ceilings on a single project.
• Not for content generation that violates Google's AUP.

Recommended configuration

import httpx

PROXY = "http://USER:PASS@gateway.squadproxy.com:7777"

def eval_gemini(prompt: str, country: str, model: str = "gemini-2-0-pro"):
    url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={API_KEY}"
    return httpx.post(
        url,
        json={"contents": [{"parts": [{"text": prompt}]}]},
        headers={
            "X-Squad-Class": "residential",
            "X-Squad-Country": country,
            "X-Squad-Session": "per-request",
        },
        proxies=PROXY,
        timeout=120,
    ).json()

For Vertex AI endpoints (*.aiplatform.googleapis.com), the same proxy shape applies with the Vertex auth headers.

Gemini-specific eval notes

• Image and multimodal eval — Gemini's image understanding and generation capabilities are evaluated similarly, with the origin anchored to match the test's regional hypothesis.
• Long-context evaluation — Gemini's long-context windows (1M+ tokens) mean individual calls are heavier. Plan concurrency accordingly.
• Regional model variants — occasionally Google deploys region-specific variants. Eval from multiple origins surfaces these if they exist.

Plans that fit

See pricing. Team plan covers most continuous Gemini eval work; Lab plan covers larger APAC-focused fleets where the JP / KR / SG origins are primary.

Related

• LLM evaluation use case
• Multilingual benchmark post
• ChatGPT API landing
• Claude API landing