Migrate from OpenRouter

​

The one-line change

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.nano-gpt.com/api/v1",  # was https://openrouter.ai/api/v1
    api_key=os.environ["NANOGPT_API_KEY"],       # was OPENROUTER_API_KEY
)

reply = client.chat.completions.create(
    model="openai/gpt-6.1-sol",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.nano-gpt.com/api/v1", // was https://openrouter.ai/api/v1
  apiKey: process.env.NANOGPT_API_KEY,        // was OPENROUTER_API_KEY
});

const reply = await client.chat.completions.create({
  model: "openai/gpt-6.1-sol",
  messages: [{ role: "user", content: "Hello!" }],
});
console.log(reply.choices[0].message.content);

curl https://api.nano-gpt.com/api/v1/chat/completions \
  -H "Authorization: Bearer $NANOGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/gpt-6.1-sol", "messages": [{"role": "user", "content": "Hello!"}]}'

​

Check these four things first

1. Your model IDs. All 10 models in OpenRouter’s weekly top ten on November 7 had the same ID here. About half of its full catalog matched exactly, and some others still resolve. See Model IDs.
2. Your provider settings. The object works, but selecting a provider adds a markup, normally 5%, and provider IDs can differ. See Routing.
3. How you read cost. usage.cost can be missing, and it isn’t always in USD. See Cost and usage.
4. Your prices. NanoGPT has no percentage fee on card top-ups, though tax can apply at checkout, and many popular models have the same list price on both services. But prices differ by model, in both directions, and some models cost more here even after OpenRouter’s fee. Compare the models you use on API pricing before you move traffic.

​

Mapping table

​

Model IDs

curl -s https://api.nano-gpt.com/api/v1/models | jq -r '.data[].id' | grep -Fx 'anthropic/claude-opus-5.5'

​

Routing

• Selecting a provider costs more. With provider routing controls or a routing suffix, billing uses that provider’s price plus a markup, normally 5%, and the request is treated as pay-as-you-go. Leave provider out to pay the model’s default price. See Billing and Markup.
• Provider IDs can differ. Find a model’s IDs with GET /api/models/{id}/providers. An unknown ID in provider.only returns 400. In provider.order it’s ignored, unless allow_fallbacks is false. Then order is the full list and is enforced like only, so saved preferences can’t add to it: if none of its IDs is usable, or none is also in only, the request returns 400 instead of going to a provider you didn’t list. That also applies to models that don’t support provider selection: if their fixed route isn’t in the list, you get 400 no_allowed_provider_for_model.
• Fewer errors move to the next model. models tries the next model after a provider failure or a prompt that’s too long for that model. Authentication, balance and invalid-request errors end the request. A first model that doesn’t meet your provider limits also ends the request, while OpenRouter moves on. See Model fallbacks.

​

Cost and usage

"usage": {
  "prompt_tokens": 10,
  "completion_tokens": 2,
  "total_tokens": 12,
  "cost": 0.0000013,
  "currency": "USD"
}

• usage.currency matters. usage.cost is in USD or, if the request was paid from a Nano balance, XNO. On OpenRouter it’s always in credits, which track USD.
• usage.cost can still be missing. If billing takes longer than the wait, or the response came through an internal fallback, there may be no usage.cost. x_nanogpt_pricing then holds an estimate with isEstimate: true. Don’t treat an estimate as the charge.
• 0 is a real value. A request covered by a subscription reports 0.
• cost_details and is_byok aren’t returned on Chat Completions.
• Embeddings report tokens but no cost. Use the Usage API.

​

Reasoning

• Thinking can be off by default. Some models think by default on OpenRouter but not here. To turn it on, send reasoning: {"enabled": true} or add the :thinking suffix. See Extended Thinking.
• Invalid values aren’t always refused. An unknown reasoning.effort returns 200. On a model that always thinks, reasoning.enabled: false hides the reasoning, but the tokens are still counted.

​

Models endpoint

​

Errors

{
  "error": {
    "message": "Model example/model is not supported on /v1/chat/completions.",
    "type": "invalid_request_error",
    "param": "model",
    "code": "model_not_supported"
  }
}

​

Known differences

• Long prompts. transforms has no effect. A prompt that’s too long returns 400 context_length_exceeded, so trim it yourself or add a longer-context model to models.
• Web search. plugins has no effect, so plugins: [{"id": "web"}] runs no search and raises no error. Use :online. web_search_options does turn on search here, and search is billed.
• Stop sequences. If a stop string matches the very start of the answer, NanoGPT can return 502 empty_response, while OpenRouter returns an empty success.
• Structured output. Support depends on the model. A model without it returns 400 model_response_format_unsupported.
• Response model. Suffixes such as :floor stay in the response’s model. OpenRouter returns the base ID.
• native_finish_reason. Not returned. Use finish_reason.
• Validation. Unknown keys in provider and invalid values in some fields are accepted without an error. A 200 doesn’t prove a setting took effect.
• App attribution. HTTP-Referer and X-Title don’t create an app page or a ranking entry.