AI API aggregators: one key for many models, and the question of who runs them

How an aggregator works

Your application sends each request to one address, with one key. The aggregator reads the model field, passes the request to the company that actually runs that model — its operator — and relays the answer back to you. Most aggregators copy OpenAI's interface: the same POST /chat/completions, the same JSON, the same streaming. The official OpenAI libraries therefore work after two changes, the base URL and the key.

curl https://evanesca.app/api/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemma-4-26b", "messages": [{"role": "user", "content": "Hello"}]}'

Moving to another model means changing the value of model — from gemma-4-26b to deepseek-v4.1-flash, say. Nothing else in your code moves. That is the whole appeal, and the whole risk: the string you edit can also change the country, and the company, your data goes to.

Why teams use one

• One integration instead of several: one way to authenticate, one set of rate limits, one invoice.
• Models change without code changes: when a better or cheaper model appears, you try it by editing a string.
• Fair comparisons: the same prompts sent to several models, and the costs read on one statement.
• One place for the rules: spending caps, which models a key may call, who in a team can use what.

The price of that convenience is a middleman. The aggregator sees every request, adds its margin, and decides where each request goes. The first two show on its price list. The third deserves the closest look, because it decides which laws apply to what you send.

The real difference: who runs the model

Every request ends up on a server run by some company, and two questions follow that are often merged into one. Where does the computation happen? That is data residency. Which company controls it, and which state can give it orders? That is jurisdiction. A service can answer the first with an EU region while the second stays American.

The CLOUD Act lets US authorities order a provider under US jurisdiction to disclose data in its “possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States” (18 U.S.C. § 2713). The US Department of Justice adds that US jurisdiction “is not limited to U.S. corporations” (CLOUD Act FAQ, 2019). The order follows the company, not the data centre.

So the useful question to put to an aggregator is not “do you have an EU region?” but “for this model, which company runs it, and where is that company controlled from?” — asked model by model, because one catalogue can mix several answers. Evanesca's sovereignty page applies that test to each of its models.

Four well-known examples, read on their own pages

The table sums up what four widely used services say about themselves in their own terms and documentation, read on 6 October 2026. It is not a ranking: each is built for something specific, and three of them — Together AI, Groq and Fireworks AI — run models on their own infrastructure rather than routing to other hosts.

All four are US companies or part of a US group. An EU region changes where a request is processed; a contract signed with a UK subsidiary changes which court hears a dispute. Neither changes which government can compel the group that controls the data. That says nothing against how carefully they handle data — Fireworks AI, for one, documents zero retention by default — it is simply a different axis.

Where Evanesca fits

Evanesca is an AI agent and chat first, and the same models answer through an OpenAI-compatible API — chat completions, the list of models, embeddings and images — with the key sent as a bearer token (API documentation). Its models, from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google, are run in Europe: by Scaleway in Paris for most of them, with OVHcloud in Gravelines as a fallback, by Mistral AI in its EU region, by AKI.IO in Germany and by Infomaniak in Switzerland — all European companies. OVHcloud, Scaleway, Mistral AI and Infomaniak have no US parent, which puts their models outside the CLOUD Act's direct reach; AKI.IO does not publish its shareholders.

Evanesca serves no model outside Europe, so no key can reach one; and a model is never sent elsewhere when its operator fails: the request fails instead.

Evanesca does not store prompts or answers: they stay in memory while the answer is written, and of each call it records the model, the tokens and the cost. The operators set their own rules, as their pages said on 6 October 2026. OVHcloud states “Zero data retention: We keep only the data required for billing purposes” (OVHcloud); Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days (Scaleway).

There is no subscription: credit from €2 that never expires, and prices per million tokens with VAT included (pricing). The exact cost of each answer is shown under it.

How to choose an aggregator

1. For each model you plan to use, ask which company runs it and where that company is controlled from. A region is not an answer.
2. Read what the aggregator keeps, then what each host keeps: logging, retention and training are decided at both levels.
3. Ask what happens when a host fails. Some services reroute the request to another host, possibly in another country; others let it fail.
4. Compare prices on one basis: per million tokens, with VAT included or excluded on both sides.
5. Match the catalogue to your needs. A marketplace lists hundreds of models; a curated catalogue lists fewer, documented one by one.

If catalogue breadth, or the lowest price on a widely hosted model, comes first for you, a US marketplace will often win. If you have to answer for which government can compel access to your data, the operator decides — and you should be able to read its name before every call. The arithmetic of running your own GPU instead is in our self-hosting cost guide, and the questions to ask a provider in writing are in how to choose an aggregator for your company.

Frequently asked questions

Read next

• How to choose an AI API aggregator for your companyFive questions to ask an AI API aggregator before you sign — jurisdiction, data kept, billing, catalogue, compatibility — and how five providers answer them.
• Open-source AI gateways and inference servers you can host yourselfLiteLLM, vLLM, SGLang, Ollama, TGI, Xinference: what each self-hosted tool does, what running it costs, and when a managed EU API is the simpler route.
• OpenRouter alternatives and competitors: what each one changesOpenRouter alternatives and competitors compared on what makes teams leave: jurisdiction, data handling, fees, speed or fine-tuning. Dated and sourced.