Who Built the Model Behind Your AI Tool?
An AI website can build its own model, serve somebody else’s or combine both. Here is how to check the maker instead of trusting the landing page.
If an AI website calls a generator “our AI,” that still does not tell you who built the model, because the company may have trained it, fine-tuned an existing system, connected another developer’s API, or combined several routes inside one polished interface, and the homepage will often make all four situations look the same.
There is nothing automatically suspicious about selling access to somebody else’s model. Useful software has always depended on technology made by other companies, and a platform may add a far better interface, editing tools, storage and simpler billing. The distinction starts to matter when a site charges a large premium, describes common technology as exclusive or gets compared with another service that may be running the same model underneath.
Our guide to AI models, tools, platforms and providers explains the terminology; this page shows how to investigate a real product when the landing page would rather give you a glowing animation than the name of its model.
Research note: We checked the public pages linked in this article on 15 September 2026. This method can establish what a company publicly discloses, but it cannot reveal private routing or an undisclosed base model without technical access that ordinary users do not have.
Quick Answer
An AI product can use a first-party model, an outside model, or a mixture of both. To classify it, identify the exact model and version inside the generator, confirm the developer through official documentation, then check the tool’s terms for third-party model providers; if the public material still does not identify the underlying model, describe it as undisclosed rather than guessing.
A company has the strongest first-party case when it publishes a named model, explains its versions and capabilities, and identifies itself as the developer. If the interface or documentation names OpenAI, Google, ByteDance, Black Forest Labs or another developer, that particular feature is using an outside model even though the tool owns everything around it. A service belongs in the mixed category when different features use different arrangements.
| Evidence you find | What it supports | Confidence |
|---|---|---|
| Another developer and model are named in the selector or documentation | That feature uses an outside model | High |
| The company publishes its own named versions, model documentation and access route | Strong evidence of a first-party model | High |
| The terms confirm that some models come from third-party providers | Outside models are used somewhere, but not necessarily in every feature | Medium |
| The page says only “our AI” or “proprietary technology” | No clear conclusion about who trained the base model | Low |
| No model or developer is identified | The underlying model remains undisclosed | None |
“Powered by proprietary AI,” “our advanced engine” and similar phrases are not evidence of model ownership, because they could describe a fine-tune, a routing system, private prompting or ordinary software built around another company’s API.
Start with the exact model, not the logo
The large brand name above the generator normally identifies who designed the product and took your payment, which is useful information but does not necessarily identify who trained the model. WaveSpeed’s model catalogue, for example, describes a single platform containing more than 1,000 models and openly names providers including ByteDance, Google, Alibaba and OpenAI. Its current routes label Seedance 2.5 with ByteDance and GPT Image 2 with OpenAI, so WaveSpeed is not pretending that every engine in its catalogue was trained in-house.
“I made it with WaveSpeed” therefore names the platform, while “Seedance 2.5 through WaveSpeed” identifies the model and the route used to access it. That longer description matters when comparing quality or price, because two websites can serve the same model while offering different settings, applying their own processing or charging completely different amounts for failed generations.
Asking whether an entire website is a “wrapper” can therefore produce a lazy answer, because a large platform may use its own technology in one feature and an outside model in the next tab. The useful question is which model produced this specific output.
How to check who built the model
1. Open the generator and record the exact version
Look beside the resolution, generation mode, settings or price before searching the homepage. Many platforms identify the technology clearly inside the product even when their front page presents the service as one giant AI system, and the version matters because two releases from the same model family may behave very differently.
If the interface only gives you a label such as “Fast,” “Creative” or “Magic,” record that wording instead of trying to identify the model from its visual style, since similar output is not proof that two tools share an engine.
2. Search the documentation and API catalogue
Documentation usually has to be more precise than advertising because developers need model IDs, supported inputs and version information. Search the site for the displayed model name, or use queries such as site:example.com model, site:example.com API and the product name followed by model provider.
Model cards, API routes, version histories and deprecation notices are stronger evidence than a homepage claim, although they still need to identify who developed the model rather than merely who hosts it.
3. Find the model announcement
A proper release will normally give the model a stable name, explain what changed and show how people can access it. The announcement does not prove that every quality claim is true, but it tells you much more about ownership than an affiliate page describing a new menu option as a revolutionary model.
4. Read the terms and privacy material
Search these pages for model, provider, third party, subprocessor and API. Terms can confirm that outside model providers are involved, while privacy or subprocessor pages may show whether uploaded material can pass through another company’s systems. They may still leave individual features unnamed, so treat a general third-party clause as evidence that outside models are used somewhere rather than proof about every tool on the site.
5. Confirm the model from the developer’s side
When a smaller service says that it uses a particular OpenAI, Google or ByteDance model, look for the same name in that developer’s official documentation. The absence of a public partnership announcement does not prove that the integration is false, especially when standard API access is available, but a supposed model that exists only on the reseller’s website deserves a cautious description.
Verification rule: the cleanest result is a matching model name on both sides, supported by a version or API reference. If the tool names the model but the claimed developer does not, record the mismatch; if neither identifies the developer, the answer remains undisclosed.
We applied the method to Higgsfield
Higgsfield is a better test than an obvious model catalogue because its public site combines outside model names, Higgsfield-branded features and the sort of glossy presentation that makes the entire product feel like one piece of technology. We checked its homepage and terms on 15 September 2026, using only pages available to any visitor and without inspecting private requests from the generator.
| Check | What we found | What it supports |
|---|---|---|
| Product listings | The Higgsfield homepage lists Seedance 2.5, Nano Banana Pro, GPT Image 2.5 Sunburst and a feature powered by GPT-6 Astra | Several features visibly use technology carrying other developers’ model names |
| Higgsfield branding | The same page presents Higgsfield Genjutsu as a “new model” and gives it its own product area | Higgsfield also markets branded model technology of its own |
| Third-party clause | Section 8 of the terms says certain features and models are provided by third-party model providers | Outside model providers are explicitly part of the service |
| Model changes | Section 1.7 says Higgsfield may add, modify, substitute, retire or remove models and model versions | The model behind a feature can change while the platform remains available |
The supported conclusion is that Higgsfield is a mixed AI platform: it provides access to clearly named outside models while also offering Higgsfield-branded technology and workflows. The pages we checked do not establish that every component of Genjutsu was trained from scratch, so describing its undisclosed base architecture as fact would go beyond the evidence.
The evidence supports “Higgsfield uses outside models,” but it does not support “Higgsfield built none of its AI.” Calling the whole platform a wrapper may feel satisfying for about five seconds, but it explains less than naming the model behind the feature somebody paid to use.
Midjourney, ChatGPT and WaveSpeed belong in different categories
| Product | Official evidence checked 15 September 2026 | Classification supported by that evidence |
|---|---|---|
| Midjourney | Midjourney describes itself as a research lab building AI models, while its version documentation records releases and lets users choose between model versions | First-party model developer with its own product; named collaborations such as Niji should still be identified separately |
| ChatGPT | OpenAI says its foundation models power products including ChatGPT, and it maintains an official model catalogue | An OpenAI product using OpenAI models, although the exact model or tool used should still be recorded |
| WaveSpeed | WaveSpeed says it exposes models from many named developers through one platform, API and billing account | A platform and inference provider for many outside models; classify each route by its displayed model |
| Higgsfield | The site lists outside models and Higgsfield-branded technology, while its terms confirm third-party model providers | A mixed platform whose individual features need separate classification |
These labels describe what the official pages support, not every part of each company’s private technical stack. They should also be rechecked rather than copied forever, because model catalogues, partnerships and default routes change quickly.
“Its own model” can mean several amounts of original work
Even when a company says it developed a model, the phrase can describe very different arrangements. It may have trained a foundation model, fine-tuned an existing open model for a narrower job, created a LoRA or adapter that changes a base model, or built a private system that routes prompts through several outside models before editing the response.
All of those approaches can produce a useful product, but calling them identical makes the term “own model” close to meaningless. When ownership affects your decision, ask who developed the base model, whether the feature requires an outside API, whether the company publishes a named version history or model card, and whether the model has a documented access route beyond the promotional page.
A closed company does not have to release its weights or reveal every training detail to own a private model, so missing public code is not proof that the claim is false. It does mean that an outside reader may be unable to verify how much of the system is original, and “the company describes it as proprietary, but the base model is undisclosed” is the accurate answer until better evidence appears.
Using an outside model does not make the tool pointless
“Wrapper” is normally used as an insult for a product that places a thin interface around somebody else’s API, although the label says little about whether the product saves time. A platform can add better editing, presets, project storage, reference management, batch processing, one bill for several models and a workflow that does not require people to lose an evening installing open-source nodes.
That is why I am happy to use WaveSpeed even though WaveSpeed did not train every model in its catalogue. Moving between image and video models without maintaining separate accounts or installing each workflow locally has real value, provided that the platform names what I am using and the price still makes sense.
The criticism becomes deserved when a tool hides the underlying model, adds almost nothing, charges a ridiculous premium and encourages customers to believe that a common API result came from exclusive technology. A glamorous interface is part of a product, but it is not model documentation.
Why the answer matters before you pay
Knowing who built and serves the model tells you whether two subscriptions provide access to genuinely different technology, whether another provider may offer similar output for less money and which company’s usage rules could apply. It also matters for privacy, because a file uploaded to one platform may need to be processed by a separate model provider depending on how the service is built.
Reviews need the same detail, which is why our TGK model-testing method records the model, version, provider, settings and test date instead of expecting a brand name to make the result reproducible. Our AI video generator comparison also names where each model was run rather than treating every interface as a new engine.
Provider terms may also allow a model to change without the product disappearing. The button can stay in the same place and the subscription can keep the same price while the output begins behaving differently, which is why a dated version number is more useful than assuming a familiar interface guarantees the same system underneath.
How to report what you found
Use one sentence that separates the model, its developer and the service that gave you access:
[Feature] used [model and version], developed by [model maker], through [platform or provider], checked [date].
When the base model is hidden, do not force the sentence to sound complete:
[Company] provided the interface, but its public documentation did not identify the underlying model when checked on [date].
That wording is useful for reviews, procurement notes and AI-generated answers because somebody else can see exactly what was verified, what remains unknown and when the information may need checking again.
Final verdict
To tell whether an AI tool uses its own model, take the exact model name from the generator, confirm the developer in official documentation, check the service’s terms for outside providers and keep the conclusion limited to the feature you investigated. The answer will normally be first-party, outside, mixed or undisclosed, and “undisclosed” is perfectly acceptable when the company has not published enough evidence.
An outside model does not make a tool useless, while a private model does not automatically make one good. The company needs to explain what is producing the result, add enough context to justify the price, and provide enough information for customers to compare the same technology elsewhere.
Sources
- WaveSpeed AI model catalogue — checked 15 September 2026
- Higgsfield homepage and model listings — checked 15 September 2026
- Higgsfield Terms of Use Agreement — checked 15 September 2026
- Midjourney About — checked 15 September 2026
- Midjourney model-version documentation — checked 15 September 2026
- OpenAI: How ChatGPT and our foundation models are developed — checked 15 September 2026
- OpenAI model catalogue — checked 15 September 2026
