Skip to content

Can API models be secretly swapped ​

Yes. You pay for a specific model, but the provider can quietly swap in a cheaper version while keeping the same price. Researchers have audited real commercial APIs with statistical methods and indeed found endpoints whose outputs did not match the official models.

What you will be able to do after this lesson ​

  • Name the several forms of model swapping
  • Understand why it is hard to detect just by looking at the answers
  • Know which situations call for extra vigilance

What forms does model swapping take ​

Two papers describe the same class of problems.

UC Berkeley's "Are You Getting What You Pay For?" (arXiv 2504.04715): users pay for a specific model, but there is no guarantee that the provider delivers it faithfully. The provider may swap in a cheaper substitute, such as a quantized version (compressing the model parameters to lower precision to save GPU memory and money) or a smaller model, while keeping the original price.

Stanford's "Model Equality Testing" (arXiv 2410.20247): to cut costs or add features, API providers may quantize, add watermarks to, or fine-tune the underlying model, changing its outputs, and they do not necessarily tell users.

Swapping methodWhat the provider gains
Swap in a quantized versionSave GPU memory and compute
Swap in a smaller modelLower costs
Fine-tune or add a watermarkAdd features, or change the model's behavior

What real audits found ​

"Model Equality Testing" used their method to audit commercial inference APIs that offered 4 Llama models in the summer of 2024: of 31 endpoints, 11 had output distributions that differed from the reference weights released by Meta.

This does not mean all 11 of these were deliberate bait-and-switches; deployment choices such as quantization can also cause differences. But it shows that the same model name may not be the same thing across different providers.

Quiz timeAn audit finds an endpoint's output doesn't match the official weights. Can you conclude it was deliberately swapped?

  • I understand that "the same model name" does not mean the same model

A real case: passing off DeepSeek as Claude ​

What was described above involves deployment differences. What follows is a more direct situation: you pay for Claude, but what you get is a different model.

The "cutting corners" documented by researchers ​

Zilan Qian, a researcher at the Oxford China Policy Lab, investigated domestic "relay services" that resell Claude. In an article published on ChinaTalk on 2026-5-5, she wrote:

  • Relay services let developers use Claude at prices as low as 10% of the official rate; a common price is 1 RMB per 1 USD of credit, which is 70%–90% cheaper than the official price
  • A user selects Opus 4.7, but the proxy can quietly route the request to Sonnet or Haiku, and in the worst case to GLM or Qwen, then label the output as the original model
  • Because the proxy sits in the middle, users cannot verify which model the request actually went to. It often only shows up on complex tasks, which is what people call "dumbing down"

On 2026-8-16, she put it more bluntly on the High Capacity podcast. Relay services call model swapping "cutting corners": the mild form is labeling it Opus 4.8 while actually serving Opus 4.6; the aggressive ones swap models outright, "for example, passing off DeepSeek as Claude, or Doubao as GPT" (original words: "swap, for example, DeepSeek for Claude or Doubao for GPT"). She said this is sometimes easy to spot, and many users on forums complain that "the model suddenly got dumber."

Why this is so easy to do ​

DeepSeek officially provides an interface compatible with the Anthropic format, at the address https://api.deepseek.com/anthropic. The official documentation states: model names starting with claude will also be accepted, claude-opus* maps to deepseek-v4-pro, and other claude* map to deepseek-flash.

This is a legitimate DeepSeek feature that makes it convenient for users to switch to DeepSeek within tools in the Claude ecosystem; DeepSeek itself is not impersonating anyone. The problem lies with middlemen: a relay service only needs to forward your request as-is to this address. You requested claude-opus-…, the response format looks just like Claude's, and you don't need to change a single line of your code — yet the backend is DeepSeek. A matching request format and model name cannot prove that Claude is behind it.

Try itHow a relay swaps Claude for DeepSeekIllustrative animation

Your code

model: "claude-opus-…"

API relay

Forwarded as-is

Anthropic official

api.anthropic.comReal Claude

The answer you received

{ "type": "message", "model": "claude-opus-…", "content": [ … ] }

Your code sends the request with the model name claude-opus-…

Quiz timeThe relay's responses have Claude's format and model field. Can you be sure Claude is behind it?

What academic audits found ​

The paper "Real Money, Fake Models" from Germany's CISPA (arXiv 2603.01919) audited "shadow APIs," that is, third-party services claiming to offer indirect access to official models:

  • Found 17 shadow APIs that had been used in 187 academic papers, and audited 3 of them in detail
  • Compared with the official APIs, the performance gap reached up to 47.21%; 45.83% of fingerprint tests failed to verify the claimed model identity
  • Fingerprint comparisons showed: endpoints labeled GPT-5 at two shadow APIs behaved like GLM-4 or DeepSeek-V3; requests labeled as the thinking-mode DeepSeek-Reasoner performed like the non-thinking DeepSeek-Chat
  • By the time the paper was written, 4 of the 17 had already ceased operations

This paper audited GPT, Gemini, and DeepSeek, and did not test Claude. But it shows the same thing: the label says an expensive model, while the fingerprint points to a cheap one.

  • I understand why "matching format and model name" does not mean genuine

Why it is hard to detect ​

  • Asking it who it is does not help. The Berkeley paper tested this: models usually only state their family or developer, and rarely their parameter count or quantization status; when asked directly about specifications, the success rate is below 5%. When a system prompt tells a model to impersonate another identity, most models fail to recognize themselves
  • Quantized versions score almost identically. The differences between quantized versions and the originals on MMLU, GSM8K, MATH, and GPQA Diamond are usually within one standard deviation of repeated runs, hidden by natural fluctuation
  • The same model can also differ. Different providers' inference settings can make the exact same model produce different outputs, which is easy to misjudge

When you should be on guard ​

  • Prices are clearly below the official rate, for example only 10–20% of the official price
  • Fine on simple questions, but inconsistent on complex tasks
  • The response style and length differ from what you see on official channels
  • The provider does not state whether it deploys the original or a quantized version

Checkpoints ​

  • Swapping methods: quantization, swapping in a smaller model, fine-tuning, watermarking
  • Audit example: 11 of 31 endpoints did not match the official weights
  • Real cases: researchers documented relay services passing off DeepSeek as Claude; the compatible interface makes this substitution possible without changing a single line of code
  • Asking about identity or running a benchmark once makes subtle substitutions hard to detect

Common Misconceptions ​

"I asked it 'which model are you,' and it said X, so it must be X." A model's self-reported identity is unreliable—change the system prompt and it will say whatever you tell it.

"The MMLU score matches after a run, so it's genuine." Small changes like quantization are almost invisible in benchmarks.

Next Section ​

So how do you actually verify it? The next section lists several methods and what each one can detect: How to Verify an API Is Genuine.

gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.