Switch from OpenRouter to Tokely

1. Check the contract you actually use

Find the exact model in Tokely's catalog. Confirm Chat Completions or Responses support, tools, streaming requirements and output limits. Text endpoints use the documented text contract. OpenRouter provider-routing objects, BYOK and unsupported endpoints cannot be assumed to work.

2. Compare your request's price

Take input/output/cache counts from a representative response and enter them in the cost calculator. Keep the model version fixed. Include your payment fees and any special units separately; a discounted input price alone does not determine the total bill.

3. Create a Tokely key and change the client

Create an account, fund its balance and create an API key in the dashboard. Set credentials through environment variables rather than putting them in source code. This example uses a supported Chat Completions model:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.tokely.me/v1",
    api_key=os.environ["TOKELY_API_KEY"],
)
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize this in one sentence."}],
    max_tokens=128,
)
print(response.choices[0].message.content)

For a Responses-only model, follow the Responses contract instead: it is stateless, requires supplied history and uses store=false.

4. Verify behavior before routing production traffic

1. Check a normal response and the recorded usage for your model.
2. Check streaming and stop behavior if your app uses them.
3. Check tool definitions/results and finish reasons if applicable.
4. Check insufficient-balance, rate-limit and malformed-request behavior.
5. Use the returned request ID to find the corresponding dashboard log.

Your application's requests can spend its funded balance. This checklist is a migration procedure, not a claim that a new paid benchmark was run.

5. Move one path and keep rollback simple

Keep the original configuration available and switch one supported request path first. Set an account spending limit and model restrictions where useful. Observe your own application's errors and costs before expanding traffic.

Common differences to account for