ocoolai is an LLM API aggregator: a service that offers access to a hundred or more AI models through one interface, at prices it advertises as below the model providers’ own, with direct connectivity for users in mainland China who cannot otherwise reach OpenAI or Anthropic. That is a real and genuinely useful product category, and several reputable companies operate in it. It is also a category where the specific operator matters enormously, because you are routing your prompts through an intermediary. On ocoolai’s public pages I could not find a named company, a terms of service link, or a privacy policy, and its domain registration lists the owner as hidden. None of that proves anything is wrong. All of it changes what you should verify first.
This article covers how the category works, why prices can legitimately sit below official rates, and the three questions to answer before you route anything through any service like this.
What the service is
ocoolai presents itself, in Chinese, as a one-stop AI model aggregation platform for unified discovery, comparison and use of models. It offers a chat interface and API integration aimed at individual developers and teams, with promotional material citing access to 100 to 200 plus models, usage-based billing, and direct domestic connectivity.
That last point is the real product. For developers inside mainland China, direct access to Western model APIs is not straightforward, and a domestic endpoint that works reliably is worth paying for on its own. Understanding that explains the category’s existence far better than the pricing claims do.
How prices can sit below official rates
The “cheaper than official” claim is the one that makes people suspicious, so it is worth explaining properly. There are several legitimate mechanisms and at least one that should worry you.
Legitimate:
- Volume commitments resold. A reseller commits to large spend, receives discounted rates, and resells at a margin that still undercuts list pricing. Standard practice across cloud and software.
- Routing to genuinely cheaper models. Chinese domestic models are substantially cheaper than Western frontier models at comparable quality for many tasks. An aggregator defaulting to those is not doing anything clever, it is passing on a real price difference.
- Batch and cache pricing. Providers offer significant discounts for cached input and batched requests. An aggregator with many users can access those tiers where an individual cannot.
- Loss-leading for growth. Venture-funded services routinely sell below cost to acquire users. Cheap now, repriced later.
The one to worry about: access obtained through accounts the provider did not authorise for resale, or through credentials the reseller does not properly own. This does happen in this category. The consequence for a customer is that access disappears without warning when the upstream account is terminated, and any prepaid balance goes with it.
You usually cannot tell which mechanism you are dealing with from outside. What you can do is ask, and treat an evasive answer as an answer.
What the public record shows
Stated factually, because these are the inputs to your own judgement rather than conclusions:
- The primary domain was registered on 19 March 2024, making it relatively young for a service handling developer credentials and payments
- Domain owner information is hidden in the registration record
- Hosting is in the United States while nameservers resolve to China
- The SSL certificate is domain-validated only, the lowest validation tier, which confirms control of the domain and nothing about the organisation behind it
- The site’s public pages name no company or legal entity, and I found no linked terms of service or privacy policy
- Promotional write-ups circulate as anonymous GitHub gists rather than through the company’s own channels or recognised publications
- Several related domains exist alongside the main one
Hidden domain registration is common and often just privacy protection. A young domain is not a red flag by itself. Anonymous gist promotion is a marketing choice. Any one of these in isolation means little. Together, for a service that will hold your API traffic and your money, they describe a profile where the sensible move is a small test rather than a commitment.
Three questions to answer before using any API reseller
1. Where does your prompt data go, and who can read it? This is the one that matters most and it is not about money. Every request you send passes through the intermediary in plain text before reaching the model. If you have no privacy policy to read and no named entity to hold responsible, you are trusting an unknown party with whatever you send. For hobby projects that may be acceptable. For customer data, proprietary code, or anything under a confidentiality obligation, it is not, and no price saving compensates.
2. What happens to your balance if upstream access stops? Prepaid credit with an intermediary is an unsecured claim on that intermediary. Ask what the refund policy is, get the answer in writing, and never prepay more than you would be comfortable writing off.
3. Does using it put your own accounts at risk? Model providers’ terms generally restrict unauthorised resale of API access. If a reseller is operating outside those terms, the exposure sits with them rather than you, but a service disappearing mid-project is a real operational risk if you have built on it. Keep your integration portable.
The safer route to the same saving
If your motivation is cost rather than connectivity, you can get most of the saving without an unidentified intermediary, because the cheap models are cheap at source.
Open-weight models have closed most of the quality gap at a fraction of frontier pricing, and you can reach them through established aggregators with published legal entities, or direct from the model providers’ own platforms. Our comparison of GLM 5.1 vs Claude Opus 4.6 covers the substitution question, Qwen3 235B vs GPT-OSS 120B compares two open options directly, and K2 Think covers another. For the broader decision, our roundup of the best AI for coding is the practical starting point.
If your motivation is connectivity from inside mainland China, that is a genuine constraint and aggregators solve a real problem. The advice then is not to avoid the category, it is to pick an operator that tells you who it is, and to keep sensitive material out of it regardless.
The bottom line
LLM API aggregation is a legitimate category solving a real problem, particularly for developers who cannot reach Western APIs directly. ocoolai operates in that category and I have no evidence of wrongdoing by it.
What I can say is that a service with no named entity, no visible terms or privacy policy, and hidden registration is one you should test with a small balance and non-sensitive prompts before relying on it, and one you should keep your code portable away from. That is not an accusation. It is the standard you would apply to any supplier you cannot identify.
