Is your “Claude” or GPT API real? How to detect relay model swapping

What “swapping” and “downgrading” mean#

• Model swap: you ask for a frontier model, the relay serves a cheaper one.
• Quantization/downgrade: the same family, but a lower-precision or distilled variant that reasons worse.
• Context truncation: your long prompt is silently cut, so recall drops.
• Silent drift: it's fine at signup, then degrades once you've committed.

Why “what model are you?” doesn't work#

Asking the model to identify itself is trivially spoofed — a relay can inject any system prompt or rewrite the answer. You can't detect a swap by asking; you detect it by testing capability and comparing against a known-good reference.

The method#

1. Fix a probe set. Multi-step reasoning, long-context needle recall, strict JSON formatting, and refusal behavior.
2. Pin parameters. temperature=0, same max_tokens, same system prompt — remove randomness.
3. Measure. p50/p95 latency and error rate over several runs.
4. Diff against the official API. Same probes, same params, same day.
5. Re-run weekly. Silent downgrades happen over time, not on day one.

# Capability probe, not "what model are you?" (that is trivially spoofed).
# Pin temperature=0 and diff the output against the official API.
from openai import OpenAI

client = OpenAI(base_url="https://daoxe.com/v1", api_key="YOUR_DAOXE_KEY")
probe = "Return only JSON: {\"sum\": <2147483647 + 1>}"
r = client.chat.completions.create(
    model="YOUR_EXACT_MODEL_ID",
    temperature=0,
    max_tokens=32,
    messages=[{"role": "user", "content": probe}],
)
print(r.choices[0].message.content)  # compare across providers / over time

Red-flag checklist#

• Output quality on hard reasoning drops versus the official API at temperature=0.
• Long-context recall fails where the official model succeeds (possible truncation).
• Token accounting looks implausibly low for the output produced.
• Latency and behavior change suddenly weeks after signup.
• The provider is evasive when you ask to benchmark it.

How DaoXE is built to pass this test#

• An open, reproducible benchmark you can point at DaoXE and the official API.
• Account-scoped /v1/models — the list you see is the list you can call.
• Native Anthropic Messages, so Claude behaves like Claude (tool use, streaming).
• Transparent, per-model usage you can reconcile.

Frequently asked questions#