Cheap GPT Is Fake? Investigating Model Impersonation in Shadow APIs
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
If you heard that "GPT-5 is available for half the official price," wouldn't you be tempted to try it? Yet cases have been found where, even though the screen displays GPT-5, a different, smaller AI was actually generating the responses behind the scenes. This article summarizes research findings and highlights what we should watch out for.
ITmedia NEWS covered this in an explanatory article on November 19, 2026, reporting on research from CISPA in Germany and the University of Waterloo in Canada, among others.
The paper is titled "Real Money, Fake Models." It is publicly available on arXiv and is noted as having been accepted at the ACM CCS 2026 international conference on computer security.
The subject of the research is shadow APIs. These are unofficial services that act as intermediaries — via APIs (the interfaces that connect software systems) — for AI from OpenAI, Google, and others, claiming to offer "the same models as the official service at a lower price."
The research team identified 17 shadow API services by examining accepted papers from international conferences and searching GitHub. These services were collectively used in 187 academic papers.
Official APIs tend to be expensive and require payment methods such as credit cards. On top of that, some official APIs are unavailable in countries like China and Russia.
Shadow APIs fill that gap, offering lower prices and no regional restrictions. In fact, roughly 80% of users were affiliated with institutions in China, where official API access is restricted.
Imagine a graduate student with a limited research budget and no access to an overseas credit card. Finding a cheaper service and jumping at it would be a natural response.
The scale of adoption is also significant: the most popular service had 5,966 paper citations and 58,639 GitHub stars on its associated repository as of December 6, 2025.
Many of the papers involved passed peer review — 116 out of 187, or 62%, were accepted at international conferences and journals such as ACL, CVPR, and ICLR.
The research team created "fingerprints" from distinctive patterns in AI responses, then combined them with statistical tests to identify what models were actually running underneath.
The result: 11 of the 24 endpoints examined (45.83%) were determined to be different from the model that was advertised.
For example, responses labeled as GPT-5 or GPT-4o-mini matched the fingerprints of different AIs such as GLM-4-9B or Qwen2.5-7B — both smaller models, entirely separate from the official GPT.
There were also cases where a request for DeepSeek-Reasoner behaved like DeepSeek-Chat, without the deliberate "reasoning mode" that makes DeepSeek-Reasoner distinctive.
This appears to be substitution with cheaper models, though the paper alone cannot confirm whether it was intentional.
The research team compared models from OpenAI, Google, and DeepSeek across three popular services. They evaluated 8 models on math and science tasks, and one model from each provider on medical, legal, and safety benchmarks.
For questions that don't require deep reasoning, scores close to the official results were observed. A quick test alone would not reveal any difference.
The gap appeared when reasoning ability was required. On the AIME 2025 math exam, one shadow API showed Gemini-2.5-pro dropping by 40.00 points and DeepSeek-Reasoner dropping by 38.89 points.
On the MedQA medical exam, Gemini-2.5-flash's accuracy rate was around 37% across all three services, compared to 83.82% on the official API. On legal exams, the same model dropped by approximately 40 points.
According to the paper, the gap reached a maximum of 47.21%. The paper itself and the ITmedia article report that even when Gemini-2.5-flash's fingerprint matched what was advertised, there were cases of significant performance degradation.
Imagine a company building a medical consultation app that serves API responses directly to patients after choosing the cheapest option. An accuracy rate of less than half is not something to laugh off.
There were also differences in the ability to refuse harmful requests. In some cases, vulnerability to jailbreaking (attacks that bypass an AI's safety mechanisms) was five times worse than the official model; in others, it was lower. The direction of the deviation was not predictable.
Of the 17 services surveyed, 15 were operated by individuals whose identities were opaque. Four services have already shut down.
This means there is a real risk of sudden unavailability or of being unable to reproduce previously reported results.
Cost isn't necessarily lower either. The paper's price comparisons found some services at 0.73 times the official cost, while others were more than five times as expensive. One service sold GPT-5 at the same price as the official API, yet the volume of output returned on the GPQA benchmark was only 38% of what the official API provided.
It should be noted that this data reflects a single snapshot from September to December 2025. The paper also acknowledges that shadow APIs can change without notice.
Here is a summary comparing official APIs with shadow APIs based on what has been covered so far.
The paper's authors state that shadow APIs should not be used in research contexts. They also call on official providers to ease regional restrictions and introduce academic pricing.
The survey found that the majority of users were affiliated with institutions in China. Since official APIs are accessible from Japan, the same circumstances do not apply.
That said, it is not entirely someone else's problem. Independent developers and startups looking to cut costs may well turn to cheap APIs that claim to offer "the same as GPT."
There is also a possibility that researchers in Japan could unknowingly cite papers whose results were obtained using shadow APIs.
For businesses, the destination of any customer data entered is also a concern. Sending internal documents to a service with an unidentified operator is something best avoided.
Overseas commentary articles suggest building a simple DIY "fingerprint test." According to this article on DEV Community, the approach involves sending fixed questions with consistent settings and comparing the recorded responses against official responses.
However, this is a suggestion from a personal blog, not a claim made in the paper itself. It is safest to treat it as a rough indicator — a match doesn't guarantee authenticity, but a discrepancy is a red flag.
The paper's own proposals are somewhat more rigorous. They include using 24 or more probe questions from a fingerprinting tool called LLMmap, conducting statistical tests with 500 or more samples, and running tests at least three times while recording the endpoint URL, model version, access date, and pricing.
For individuals, three practical first steps are:
Q. Are shadow APIs illegal?
A. This article and the paper do not make a determination on legality. What the paper identifies are suspected model impersonation, performance variability, and operator opacity.
Q. Are all cheap APIs fake?
A. Not necessarily. Of the 24 endpoints examined, 11 were determined to differ from what was advertised. However, the paper also separately reports fingerprint discrepancies in 12.50% of cases and urges careful verification.
Q. Why can't you detect substitution with simple questions?
A. Because even smaller models can answer straightforward questions fairly accurately. Differences emerge with questions that require deeper reasoning, such as those in mathematics or medicine.
Q. What should businesses in Japan watch out for?
A. Prioritize verifying the operator's identity and where your data goes over the appeal of lower pricing. It is safer to choose providers with clearly established contractual relationships.
Start by taking a moment to verify the operator behind any API you are currently using, as well as the quality of the responses you are receiving.
This article is a cross-post from AI Friends.