ModelVerify

• 1. overview (549 chars)

  ModelVerify.ai is an "Anonymous Own Key (BYOK) model verification platform" for developers and enterprises. Its core mission is to provide independent evidence for the "trust issue" of AI APIs. Users only need to fill in the third-party API endpoint (Base URL), model ID and their own API Key, and the platform will compare the behavior of the endpoint with the "fingerprint" of the official model through diagnostic probes to determine whether it is the model claimed. At a time when API brokers and shell models are rampant, this tool solves a real and high-frequency pain point: the "Claude Opus" you pay to call may be running the cheaper Haiku behind it.

• 2. Product background (1012 chars)

  The official positioning of ModelVerify.ai is "Bringing transparency to AI APIs — independent model verification for everyone." It points the finger at a dark industry chain that has been ignored for a long time - third-party API transfer stations and model shelling fraud. In March 2026, Germany's CISPA Helmholtz Information Security Center released a research report "Real Money, Fake Models: Deceptive Model Claims in Shadow APIs", which found that up to 45.83% of shadow API endpoints in test samples suffered from model replacement fraud, that is, cheap open source models were used to pretend to be paid top models, which even resulted in distortion of the evaluation data of 187 academic papers. It is against this background that ModelVerify.ai transforms "model verification" from a research topic into a public evidence layer that anyone can use for free. The platform currently does not disclose parent company or financing information, and its team background has not been disclosed. However, its public evidence layer has completed more than 8,838 endpoint verifications, 443 new ones in the past 30 days, and a total of more than 1,586 "claimed inconsistent" endpoints have been found.

• 3. Product functions and usage experience (1559 chars)

  The verification process is light enough: users visit the official website and enter the Base URL, Model ID and API Key to start, without registering an account. The platform sends a set of diagnostic probes to the target endpoint, compares the returned behavior pattern with the "known-good fingerprint" collected from the real official endpoint, and finally gives the judgment (Match / Mismatch), compatibility score (0 to 1, the closer to 1, the more like the official model), risk level, stability index and latency data. For example, in the recent batch of verifications displayed on the homepage, api.vilao.ai's claude-opus-4-8 compatibility is only 0.67, and is judged to be a high-risk mismatch, while vip.j3gb.com's gpt-5.5 compatibility is 0.80, stability is 100%, and it is a low-risk match.   In addition to single verification, the platform also provides two resident capabilities. One is the "Official Supplier Drift Monitor (Model Drift Monitor)": official fingerprints of OpenAI, Azure OpenAI, Anthropic, and AWS Bedrock are collected every 6 hours and compared with the baseline. The status is marked in five colors: green (normal), yellow (observation), orange (suspected drift), red (confirmed drift), and gray (stale data). It is worth noting that OpenAI and Azure's GPT, Anthropic and Bedrock's Claude are tracked separately to avoid cloud vendor routing changes from contaminating the official model baseline - for example, monitoring shows that GPT-5.5/OpenAI has a drift score of 1.3 (observed), while Claude Opus 4.8/Anthropic is 0.7 (normal). The second is the "Official Model Benchmark List (baseline_150)", which uses scores from 0 to 1 to rank mainstream models in the four dimensions of reasoning, coding, tool calling, and latency. It is only used as a reference for capabilities rather than a strict benchmark.

• 4. Pricing and business model (777 chars)

  ModelVerify.ai is currently completely free, anonymous, and requires no account. Its business model has not yet been made public, but in terms of product form, it is more like a "public infrastructure/brand trust tool" rather than a directly monetized SaaS: users bring their own keys (BYOK), the platform does not charge API call fees, and relies on the accumulated public evidence layer to establish industry reference value. The platform's handling of API Key also reflects anonymous positioning: the key is only temporarily stored in Redis and is only used for the current run. It is deleted immediately after the run is completed and is never written to PostgreSQL or file storage. The front end never displays the decrypted key. The verification report is saved and accessed through the "report code". The anonymous test report can only be accessed by this code before expiration, and is not bound to a permanent account.

• 5. User reviews (765 chars)

  Since the product launch is relatively close to the release of documents (the documents are concentrated in June 2026), the public channel has not yet accumulated a large number of independent user long reviews, and no clear Product Hunt release page has been found. The reference "user signals" mainly come from its public evidence layer itself: 443 new verifications have been added in the past 30 days, and more than 1,586 inconsistent endpoints have been found in total, indicating that a considerable number of developers and transfer station users have used it to verify before actual purchases. From the perspective of interaction design, it addresses the core anxiety of transit API buyers who are "fearful of buying fake models". The experience of low threshold and instant results (most verifications are completed in a few seconds to more than 20 seconds) is in line with the habits of developer tools.

• 6. Industry evaluation (653 chars)

  The industry’s demand for “model validation” is rising rapidly. CISPA’s research turned “shadow API model replacement” from a rumor in the circle into a security issue supported by data. Domestic and foreign media (Tencent News, Zhihu, Sohu, etc.) concentrated on reporting on the “real money to buy fake models” incident in March 2026. Tools with similar ideas have also emerged in the open source community, such as the llm-verify project on GitHub, which aims to detect fake AI APIs. The uniqueness of ModelVerify.ai is that it makes fingerprint comparison a "continuously updated public evidence layer" instead of a one-time local script, and superimposes official model drift monitoring, which makes it more like an "industry benchmark provider" among similar tools.

• 7. Controversies and Risks (719 chars)

  To be clear, ModelVerify.ai’s determination is “evidence” rather than “judgment.” Officials have repeatedly emphasized that the determination should be based on delay, stability, and drift signals. A single "mismatch" does not necessarily mean fraud, but may also be a model version difference or a configuration issue. Secondly, fingerprint coverage determines the verification accuracy. For unpopular models that do not have official fingerprints on the platform, the results have limited reference value. In addition, as an anonymous platform, it does not provide compliance endorsement from suppliers, and users still need to make decisions based on their own testing and supplier reputation. Finally, the platform team and operating entities are not disclosed, and there is uncertainty about long-term availability and data policies.

• 8. Applicable groups and usage suggestions (862 chars)

  This tool is most worthy of use by AI application developers, API transfer station purchasers, and researchers doing model evaluation, especially those teams that route traffic to third-party LLM APIs or gateways but cannot confirm the real models behind them. The usage advice is very straightforward: before cutting production traffic to a certain endpoint, run a verification with it first, focusing on whether the compatibility score is close to 1, whether the risk is low, and whether the stability is 100%; at the same time, pay attention to the official drift monitoring to confirm that the official model you rely on has not experienced any recent degradation in capability. It is not suitable to regard it as "the entire basis for supplier credit endorsement" - it answers "does this endpoint now behave like the declared model", not "whether this company can be trusted". Alternatives include self-built probe comparison, open source llm-verify, and similar Chinese verification tools such as Ofox and API-CHECK.

• 9. Summarize (306 chars)

  ModelVerify.ai has turned a long-undervalued trust vulnerability into a practical tool for free, anonymous, and reusable public evidence layers. It is a link worth adding to the "pre-launch checklist" in the AI ​​API procurement and evaluation link; its value will continue to amplify with the accumulation of fingerprint coverage and public evidence layers.

No in-depth report yet

• 

  LIsul_x

  —

  客观说一句，别把它当尚方宝剑。它给的是「行为像不像宣称模型」的证据，不是供应商信用背书。我一般会结合三点：兼容性分数接近 1、稳定性 100%、再去看官方漂移监测确认模型本身没退化，三样齐全才放心切流量。冷门模型没指纹的，它也没辙，这时候还是得自己写探针兜底。

---

• 

  JTorres_Pro559

  —

  API Key 只暂存 Redis、跑完就删，这点比那些要你注册绑定账号的检测站放心。不过官方漂移监测我更看重，GPT-5.5/OpenAI 最近漂移分 1.3 进观察态，说明连官方自身都在变，验真不是一劳永逸。

---

• 

  TJohnson_Max

  —

  免费的，先冲了。

---

• 

  熊丹怡

  —

  报告码能留着复盘，挺贴心的设计。

---

• 

  angrybear295

  —

  做科研评测的注意了，有论文因为用了假 API 数据直接失真。我现在凡是接第三方端点，先丢进 ModelVerify 验一遍，兼容性、风险、稳定性三项都绿才敢用，比盲信商家 description 靠谱一万倍。

---

• 

  桃花753

  —

  终于有个能验真模型的了。

---

• 

  brownbutterfly366

  —

  匿名还不用注册，这点很对我胃口，怕留痕迹。

---

• 

  TOada

  —

  研究说市面上 45% 以上的中转存在偷梁换柱，我本来半信半疑，自己拿 ModelVerify 跑了十几个端点才信。最离谱的一个声称 gpt-5.5，兼容性 0.06，稳定性还只有 33%，明显是拿别的东西顶包。这种工具早该有了。

---

• 

  星辰_7

  —

  跟 API-CHECK、Ofox 比，它胜在公共证据层，能看到别人验过的端点，省得每家都自己踩一遍。

---

• 

  Jeffrey_RichardsonZ

  —

  说真的，光看商家宣传根本分不清真假，延迟、稳定性、漂移一起看才有意义。单次不匹配不一定是欺诈，也可能是版本差异，别一棍子打死。

---

• 

  xNicolasGutiérrez_dev

  —

  拿它测了手头五个中转，三个都是不匹配，其中一个 gpt-5.5 兼容性只有 0.06，基本就是拿开源模型糊弄人。以后买额度前必验。

---

• 

  MBennettSr

  —

  兼容性分数 0.67 直接判高风险，比我肉眼比对准多了。

---

• 

  iMayaFleury_dev

  —

  团队做 AI 编程助手，高频调 Claude，官方价扛不住只能走中转。以前全靠运气，现在上线前先拿它跑一遍，兼容性低于 0.9 的坚决不上生产。

---

• 

  夏轩

  —

  之前用某中转，跑出来居然是 Haiku 顶着 Opus 的名，血亏。

---

• 

  Sofia223

  —

  踩过坑才懂这东西的价值。去年团队图便宜买了个 Claude 中转，跑代码老降智，排查两周才发现后端根本不是 Opus，而是拿便宜模型顶着名卖。要是当时有 ModelVerify，输个 Base URL 和 Key 十几秒就能看兼容性分数，也不至于白烧那么多 token 和工时。现在它已经是我采购中转前的固定一步了。

---

• 

  宋红

  —

  中转站水太深了。

No reviews yet