API providers and routers

The same model can be served by its maker, a cloud, an inference host or a router. Who serves it decides where your prompts and code go, and whose terms apply to them.

Five blank envelopes of different sizes fanned out, one orange-red

Four kinds of provider

All providers 14

Provider	Kind	Serves	OpenAI-compatible	Before you send code
OpenAI API	First-party	GPT models	Yes, native at api.openai.com/v1; Responses API recommended for new projects	Read the data policy
Anthropic API	First-party	Claude models	Test layer at api.anthropic.com/v1; native API is Messages	Anthropic says the layer is mainly for testing and not production-ready for most uses
Gemini API	First-party	Gemini models	Yes, at /v1beta/openai	Read the data policy
Grok API (xAI)	First-party	Grok models	Yes, marked legacy; Responses API preferred	Read the data policy
Mistral API	First-party	Mistral models	Same request shape, not labelled compatible	Read the data policy
DeepSeek API	First-party	DeepSeek models	Yes, base URL api.deepseek.com	Read the data policy
Amazon Bedrock	Cloud	Amazon Nova plus DeepSeek, Mistral AI, Qwen and other vendors	Yes, at /openai/v1 on bedrock-runtime	Signs in with AWS credentials or a Bedrock API key
Gemini Enterprise Agent Platform	Cloud	Gemini models in your Google Cloud project	Yes, for Gemini models	Only Google Cloud auth works with the compatible client
Microsoft Foundry	Cloud	GPT models plus Cohere, DeepSeek, Grok, Meta and Mistral models	Yes, at /openai/v1	Model availability varies by Azure region
Groq	Inference host	GPT OSS, Llama, Qwen3.8 27B and others	Mostly; some features unsupported	Read the data policy
Together AI	Inference host	DeepSeek, Kimi, MiniMax, Qwen and others	Yes, at api.together.ai/v1	Read the data policy
Fireworks AI	Inference host	DeepSeek, GLM, Kimi, Qwen and others	Yes, at /inference/v1	Read the data policy
OpenRouter	Router	465 models listed on 15 October 2026, served by many hosts	Yes, at openrouter.ai/api/v1	Sends requests to third-party hosts; set data_collection to deny to skip hosts that may store data
LiteLLM	Router	Self-hosted gateway in front of 100+ provider APIs	Yes, its default format	You run it; data goes to whichever providers you configure

Endpoints, base URLs and model lists checked against each provider’s own documentation on 15 October 2026. We do not summarise data-retention terms; read each provider’s current policy yourself before you send private code.

Switching providers is one line

Thirteen of the fourteen providers above document an OpenAI-compatible Chat Completions endpoint (Anthropic’s is a compatibility layer it describes as meant for testing), and Mistral documents the same request shape without calling it compatible. Change the base URL, the key and the model name; keep the rest. That makes it cheap to move when prices or rankings shift, and it means one review of the calling code covers every provider you switch between.

client.js JavaScript · 24 lines Reviewed


      1// Base URL and model as each provider documents them (checked 2026-10-15):
2//   Groq        https://api.groq.com/openai/v1                        qwen/qwen3.8-27b
3//   Together    https://api.together.ai/v1                            MiniMaxAI/MiniMax-M3
4//   DeepSeek    https://api.deepseek.com                              deepseek-flash
5//   Gemini API  https://generativelanguage.googleapis.com/v1beta/openai  gemini-3.8-flash
6//   OpenRouter  https://openrouter.ai/api/v1                          qwen/qwen3.8-27b
7async function ask(prompt) {
8  const res = await fetch(process.env.BASE_URL + '/chat/completions', {
9    method: 'POST',
10    headers: {
11      'Content-Type': 'application/json',
12      Authorization: 'Bearer ' + process.env.API_KEY
13    },
14    body: JSON.stringify({
15      model: process.env.MODEL,
16      messages: [{ role: 'user', content: prompt }]
17    })
18  });
19  if (!res.ok) throw new Error('HTTP ' + res.status + ': ' + await res.text());
20  const data = await res.json();
21  return data.choices[0].message.content;
22}
23 
24ask('Say hello in one sentence.').then(console.log, console.error);

Editor-written sample, run in Node 22 against a local test server, not a live provider. Base URLs and model names in the comment are as each provider documented them on 15 October 2026. Keep keys in environment variables, never in the repo.