Proxies for the full OpenAI API surface

Where this differs from the ChatGPT landing

The ChatGPT API landing is focused on the Chat Completions surface specifically (the primary consumer- branded surface). This page covers the broader OpenAI API platform: Embeddings, DALL-E image generation, the Realtime API for voice applications, and the Assistants API for agent-like multi-turn interactions.

Surface-by-surface routing shape

Chat Completions

See ChatGPT API landing for the detailed configuration. Residential + per-request for regional eval; ISP

• sticky for multi-turn agent benchmarks.

Embeddings

High-volume, latency-tolerant, cost-sensitive. Datacenter is the right class. Concurrency is the binding constraint — the OpenAI Embeddings API rate limits per account, not per IP, so the proxy layer serves to distribute connection load, not rate-limit bypass.

def embed(texts: list[str]):
    return httpx.post(
        "https://api.openai.com/v1/embeddings",
        json={"model": "text-embedding-3-large", "input": texts},
        headers={
            "Authorization": f"Bearer {API_KEY}",
            "X-Squad-Class": "datacenter",
            "X-Squad-Country": "us",
        },
        proxies=PROXY,
        timeout=60,
    ).json()

DALL-E (images.generations)

Evaluation of regional content policy for image generation is similar to the Chat eval pattern. Residential per-country for authentic origin; per-request rotation.

Realtime API

Low-latency voice. Datacenter via us-east-1 is the right shape; residential adds latency that hurts voice UX measurements. The Realtime API itself has carrier-origin sensitivity in some deployments, so for specifically-mobile realtime eval, switch to 4G mobile routing — see 4G mobile.

Assistants API

Multi-turn by definition. ISP + sticky-30m is the right shape for eval that runs the same assistant across many turns.

Plans that fit

See pricing. Large-scale Embeddings workloads are typically bottlenecked on OpenAI-side rate limits, not SquadProxy bandwidth, so the Solo or Team plan usually covers the proxy side. Continuous eval fleets across multiple surfaces scale into Team or Lab.

Related

• ChatGPT API landing
• LLM evaluation use case
• Residential vs datacenter routing