One platform.
Every model we serve.

Start

A key authenticates every request. Sign in and mint one, then keep it in REKA_API_KEY. Get an API key

Or your own client

There is no Reka SDK. The API is OpenAI-compatible, so an existing OpenAI client works after two changes: the base URL and the key.

npm install openai

Quickstart

Point an HTTP client at the chat-completions endpoint and name a model from the table below.

curl https://api.reka.ai/v1/chat/completions \
  -H "Authorization: Bearer $REKA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek4-flash",
    "messages": [
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  }'

The same call from the SDK:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.reka.ai/v1",
  apiKey: process.env.REKA_API_KEY,
});

const res = await client.chat.completions.create({
  model: "deepseek4-flash",
  messages: [{ role: "user", content: "Why is the sky blue?" }],
});

console.log(res.choices[0].message.content);

Reference Streaming, every field, the base-URL swap and the error codes are on the reference.

Models

The text models below run on hardware we operate ourselves. Prices are the /v1/models feed as of 3 September 2026.

Pricing

Cache reads on DeepSeek V4 Flash are $0.007 per million tokens, half the $0.014 charged by the cheapest of the seventeen DeepSeek V4 Flash endpoints listed on OpenRouter. At a 64% cache-hit rate the $0.007 rate works out to $0.0441 per million input tokens.

The low rate is prefill we skip, not quality we drop. The fp4 and fp8 quantisations, the prefix cache and the batch scheduler each have to hold their eval scores on a fixed suite before reaching the endpoint, and a regression rolls the change back. We list the same models on OpenRouter, which measures per-endpoint throughput, latency and error rates independently.

Under the hood

The gateway checks your key, meters the tokens, decides whether you are on shared or dedicated capacity, and hands the request to vLLM on GPUs we own. Nothing rewrites the request and nothing picks a model for you. deepseek4-flash is served fp4; glm5.2 and qwen3.8-27b fp8.

There is no reseller in the middle. When something breaks it pages us, and we fix the machine it broke on.