GPT API with live pricing, Chat Completions and the Responses API

1. Get a key

   Sign up, top up from $5 and create a key in API keys. The same key calls every model, the data tools and the browser agent.

2. Set two variables

   The OpenAI SDKs, and most apps built on them, read OPENAI_BASE_URL and OPENAI_API_KEY. Set them to https://lovelyrouter.com/v1 and your key. If an app asks for an API host and appends /v1 itself, enter https://lovelyrouter.com.

3. Call a GPT model

   The request body passes through unchanged, so OpenAI's API reference applies to every parameter.

   cURLPythonNode.js

   Shell

   curl https://lovelyrouter.com/v1/chat/completions \
     -H "Authorization: Bearer $LOVELY_API_KEY" \
     -H "content-type: application/json" \
     -d '{
       "model": "gpt-6.1-sol",
       "messages": [
         {"role": "system", "content": "Answer in one sentence."},
         {"role": "user", "content": "What does a load balancer do?"}
       ]
     }'

Prices are in US dollars per million tokens, straight from the catalog. client.models.list() returns the same list with prices and needs no key. GPT models are served in the gpt-value channel group; the model marketplace shows live availability and status shows hour-by-hour availability per channel group.

Changing vendors is a change to the model field. A model called outside its native format is served by a channel that converts the request for it, so options that only one vendor defines are only certain to work in that vendor's own API. Image models answer the Images API: see image generation.

GPT models are also served at POST /v1/responses, the API Codex CLI and the Agents SDK use. Same base URL, same key; client.responses.create() works as is. Add Lovely Router to Codex as a model provider:

model = "gpt-6.1-sol"
model_provider = "lovely"

[model_providers.lovely]
name = "Lovely Router"
base_url = "https://lovelyrouter.com/v1"
env_key = "LOVELY_API_KEY"
wire_api = "responses"

Codex sends a prompt_cache_key for each session, so every turn of a session stays on the channel that holds its prompt cache. Two things to know about the Responses API here:

• Send the whole conversation. Responses are not stored: GET, DELETE and cancel on /v1/responses/{id} return 404, and previous_response_id only works while the conversation stays on one channel. Send the full history in input with "store": false, as Codex does.
• GPT models only. Other models return 404 with not served in responses format; call them with Chat Completions or their own API.

Not served: embeddings, audio, image variations, legacy Completions, Assistants, files, batches, fine-tuning and moderation. Those paths return 404.

GPT models are billed per token: uncached input, cached input at the cache-read price, and output with reasoning tokens included, each at the model's price per million tokens. GPT models have no separate cache-write charge. Every channel group bills at list price today, so each GPT model costs what OpenAI lists. The charge is computed from the usage the model reports and taken from your prepaid balance within seconds.

You top up once and spend across every product; there is no subscription. Requests rejected before they reach a model, and attempts the gateway retried elsewhere, are not charged. See pricing for everything else the key can buy.