Engineering

How to choose an AI API aggregator for your company

Choose an AI API aggregator by asking five questions, in order: which companies process your prompts and which law can reach them; what each of them keeps; how you pay and what caps the spend; whether the models you need are there; and whether your OpenAI code runs unchanged. The first question splits the market. OpenRouter, Together AI, Groq and Fireworks AI are US companies, within the CLOUD Act's reach wherever their servers are, while Evanesca, a French service, runs its models at European operators, most of them in France. The other four questions decide between the rest.

Published 6 Oct 2026

What an aggregator adds — including a company

An aggregator puts many models behind one OpenAI-compatible endpoint. Switching models becomes a change of string, not of integration, and no single model vendor can strand you by changing its prices or its terms. AI API aggregators explains how they work; this guide is about choosing one.

It also adds a company to the path of your data. If prompts can contain personal data, that company processes it on your behalf, which under the GDPR means a contract under Article 28 and a list of its own sub-processors. Evaluate it as you would any processor, not as a developer tool.

Five questions, in order

Jurisdiction: which companies process the prompts — the aggregator, the operators that run the models, their sub-processors — and which law can compel each of them?
Retention: what does each link keep, for how long, with which exceptions, and does anyone train on it?
Billing: prepaid or invoiced, automatic recharge or not, and which limits can stop a runaway agent?
Catalogue: are the models you need there, and is it clear who runs each one, and where?
Compatibility: does your OpenAI code run unchanged, including streaming, tool calls and errors?

Ask for operational facts in writing too — a status page, an incident history, a support contact. A provider that cannot show them yet should say so plainly.

Question one: who can be compelled

A US company must disclose data in its “possession, custody, or control” when US authorities order it to, wherever the data is stored (18 U.S.C. § 2713). That is a different question from GDPR compliance: a US provider can sign standard contractual clauses and a DPA, and still receive such an order — OpenAI's case shows how.

Regional offers answer where, not who. OpenRouter's in-region routing keeps prompts and completions in the EU on its Business and Enterprise plans (in-region routing, read on 6 October 2026); OpenRouter, Inc. remains a US company. So check each link's legal entity and who controls it. OVHcloud belongs to OVH Groupe, controlled by the Klaba family (shareholders); Scaleway to the iliad group (about). Both are French with entities in the United States: out of the CLOUD Act's direct reach, not immune to it.

Evanesca's chain: FLEECE AI SASU, a French company with no US parent and no US subsidiary; its models at OVHcloud in Gravelines, Scaleway in Paris, Mistral AI in its EU region, AKI.IO in Germany and Infomaniak in Switzerland; its own servers on Vercel, a US company, in its Paris region, with no copy of your prompts kept there. Evanesca serves no model outside Europe. Details on sovereignty.

Questions two and three: what is kept, and how you pay

Provider	Your data, by default	Billing
Evanesca	No copy kept by Evanesca; never used by Evanesca for training. Operators: OVHcloud keeps billing data only; Scaleway may keep requests that cause errors or look abusive, up to 14 days; Mistral AI keeps inputs and outputs 30 days to monitor abuse	Prepaid: credit from €2 that never expires; spending limits for teams
OpenRouter	Depends on the upstream host; some may train on prompts, and can be excluded	Prepaid credits, optional automatic recharge
Together AI	Zero data retention is an opt-in setting	Prepaid balance, replenished automatically under its terms
Groq	Not retained beyond the service's needs; no training without permission	Free plan; Developer plan invoiced monthly, with spend limits
Fireworks AI	Zero retention on open models by default, except its Responses API (30 days unless opted out)	Prepaid for self-serve accounts since 1 July 2026, optional automatic recharge

On billing, the rule is simple: prepaid credit caps the spend unless something refills it automatically, while an invoice in arrears leaves the risk of a looping agent or a leaked key with you.

Question four: are your models there?

Breadth and curation pull in opposite directions. OpenRouter lists hundreds of models, closed ones included; Together AI and Fireworks AI run broad open-model catalogues on their own infrastructure; Groq serves a shorter list on its own chips. Evanesca's catalogue is mostly open-weight (Mistral Large 4, in preview, has no published weights yet), and it names the operator and the place of every model before you call it — all of them in Europe.

Compare prices on the same basis: Evanesca's include VAT, most competitors publish theirs before tax. On Evanesca, Llama 3.3 70B costs €1.512 per million input tokens and the cheapest model, Gemma 4 26B, €0.168. See the models, compare and pricing.

Question five: will your code run unchanged?

All five providers offer OpenAI-compatible chat completions; the differences hide at the edges — streaming, tool calls, structured outputs, error formats, embeddings. Run your own prompts through each candidate before you move traffic. With Evanesca, the move is a base URL and a key:

Python

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://evanesca.app/api/v1",
    api_key=os.environ["EVANESCA_API_KEY"],
)

answer = client.chat.completions.create(
    model="gemma-4-26b",
    messages=[{"role": "user", "content": "List the three riskiest clauses in this contract."}],
)
print(answer.choices[0].message.content)

Evanesca's catalogue marks which models return real tool calls, measured by a call rather than read on a model card, and its embedding models run in France as well (how to choose one). The docs cover streaming, errors and images. Evanesca's API is not open yet; the code works once it is.

Which aggregator for which company

Company profile	First need	Fits
EU company handling personal or client data	Jurisdiction and a DPA	Operators under European control — for example, the ones behind Evanesca
Team that must answer a DPO or an auditor in writing	A named operator and place for every call	The same, with a sub-processor list
Start-up that needs closed frontier models	Breadth	OpenRouter
Voice agents, live copilots	Latency	Groq, if its models fit
Team that fine-tunes	Training and serving in one place	Together AI or Fireworks AI

Many companies use two providers: sensitive work on EU operators, experiments and commodity volume on the cheapest pool. With OpenAI-compatible APIs, the split is configuration, not code. Our OpenRouter alternatives compares the providers in more detail.

“Enterprise-compliant” is something you verify

No provider can certify itself compliant. Ask for the facts that let you check: the legal entity and who controls it; the operators and where they run each model; what each link keeps; an Article 28 DPA with a sub-processor list, notice of changes and a right to object; and how the service meets the AI Act's transparency rules.

What Evanesca puts in writing for teams: a signed Art. 28 DPA, the sub-processor list with notice and a right to object, SSO and roles, spending limits, VAT invoices and exports of each answer's passport. See Business, transparency and the AI Act page.

Frequently asked questions

Which AI API aggregator should a European company choose?

Start from your first constraint. If a DPO, an auditor or a client contract asks which law can reach your prompts, choose operators under European control, such as the ones behind Evanesca: Scaleway in France for most of its models, with OVHcloud as a fallback. If breadth comes first, OpenRouter; speed, Groq; fine-tuning, Together AI or Fireworks AI.

Is there a GDPR-compliant alternative to the OpenAI API?

No service can certify itself compliant: compliance depends on the contract and on how you use the service. On jurisdiction, Evanesca offers an OpenAI-compatible API whose models all run at European operators, most of them in France, with a signed Art. 28 DPA for Business workspaces. Switching is a base URL and a key.

Do AI API aggregators train on my data?

It varies. OpenRouter says some upstream providers may train on prompts, and lets you exclude them; Groq's services agreement rules out training without permission; Fireworks AI keeps no data by default on open models, its Responses API aside. Evanesca never uses your conversations for training and keeps no copy of them. OVHcloud, Scaleway, AKI.IO and Infomaniak state that they do not train on them either; Mistral AI lets its customers object to training, and Evanesca's account objects, except for what is sent to Mistral Large 4: it is a preview, and Mistral's commercial terms let it use the data of preview models for training, whatever the objection.

Prepaid or invoiced: which is safer for spend control?

Prepaid, as long as nothing refills the balance automatically: the spend stops at what you loaded. An invoice in arrears leaves the risk of a looping agent or a leaked key with you. Evanesca is prepaid, and Business workspaces add spending limits.