Continually learning inference

Inference that learns.

The world's first continually learning inference platform.

Every request you send teaches it. Within a couple thousand requests it adapts to your workload; from there accuracy keeps rising and cost per task keeps falling.

01. The problem

Static inference never gets better.

Request ten thousand costs exactly what request one did, and the model repeats the same mistakes on both.

Everything your traffic could teach it (which answers worked, which steps were wasted, where tokens were burned) is thrown away after every call.

accuracycost per taskrequest 11,00010,000same mistakes, every time
Fig. 1, static inferenceNothing compounds
“Every request you send makes the next one better and cheaper.”The idea behind CT Router
02. The loop

Every request is a lesson.

Your traffic becomes the curriculum. The platform learns continuously in the background while it serves you.

  1. 01
    ServeAnswer every request on the model you chose, at provider list price.
  2. 02
    ObserveLearn which outputs worked for your workload, from your own traffic.
  3. 03
    DistillCompress what works into cheaper, more token-efficient execution.
  4. 04
    VerifyCheck every change against your traffic before it serves you.
01 serve02 observe03 distill04 verifyyour trafficis the curriculum
Fig. 2, the learning loopServe · observe · distill · verify
03. The curve

Better in days. Cheaper in weeks.

Changes show up within a couple thousand requests, usually in the first few days. Over the following weeks cost per task keeps decreasing and accuracy keeps increasing.

It compounds: the more traffic you send, the faster it learns and the cheaper each task gets.

~2k
requests to first adaptation
days
to first visible gains
weeks
of compounding savings
first adaptation~2k requestsaccuracy ↑cost per task ↓day 1week 1week 4+illustrative, not measured data
Fig. 3, illustrative trajectoryAccuracy up · cost down
04. Drop-in

Change one line.

Point any OpenAI SDK at https://ctrouter.ai/api/v1. Learning starts with your first request; there's nothing to label, train or deploy.

from openai import OpenAI

client = OpenAI(
    base_url="https://ctrouter.ai/api/v1",
    api_key="$API_KEY",
)

completion = client.chat.completions.create(
    model="anthropic/claude-sonnet-5.5",
    messages=[{"role": "user", "content": "What is the meaning of life?"}],
    extra_headers={"HTTP-Referer": "https://your-app.com", "X-Title": "Your App"},  # optional
)
print(completion.choices[0].message.content)
import OpenAI from "openai";

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

const completion = await client.chat.completions.create({
  model: "anthropic/claude-sonnet-5.5",
  messages: [{ role: "user", content: "What is the meaning of life?" }],
});
console.log(completion.choices[0].message.content);
curl https://ctrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-5.5",
    "messages": [{"role": "user", "content": "What is the meaning of life?"}]
  }'
05. Models

Start from any frontier model.

30 models from 9 providers. Learning runs on top of whichever one you choose.

All 30 models →
06. Pricing

You pay less as it learns.

Day one is provider list price per token. As it learns, tasks take fewer tokens and run on cheaper execution, so your bill falls with no work on your side.