Advisor API: Give Your AI a Second Opinion
Most questions are fine with one model. Some decisions are important enough to deserve a second opinion.
That might be a database migration, a security review, a research conclusion, or a plan where a bad assumption would be expensive. You could send the same prompt to two models and compare the answers yourself, but then you have to decide what to share and combine the results.
NanoGPT's Advisor API handles that inside one request.
You choose a main model and a different advisor model. The main model can ask the advisor a focused question, consider its feedback, and then give you one final answer.
How it works
The main model starts with your request. Depending on the mode you choose, it can either answer directly or pause to ask the advisor for help.
If it consults the advisor:
- The main model decides what it wants reviewed.
- The advisor answers that focused question.
- The main model uses the feedback to prepare the final response.
Your app still receives one normal answer from the main model. You do not have to merge two separate responses.
A basic request
Add an advisor section to a non-streaming Chat Completions or Responses API request:
{
"model": "gpt-4o-mini",
"messages": [
{
"role": "user",
"content": "Review this migration plan and identify the riskiest assumption."
}
],
"stream": false,
"advisor": {
"model": "anthropic/claude-opus-4.8",
"mode": "auto"
}
}
The top-level model is the main model. advisor.model is the second model it can consult. They must be different models.
Auto or required?
Use auto when a second opinion might help, but is not worth using on every request. The main model can answer a simple question directly and consult the advisor for a harder one.
Use required when the review is the point of the request. This works well for approval steps, important decisions, and tests where you need to know that a consultation was attempted.
required does not mean the advisor is correct, or that the two models will agree. It means the main model must try to get the second opinion before answering.
Give the advisor a clear role
The feature works best when the advisor has a specific job.
For example:
- A fast model drafts a plan and a stronger model looks for failure cases.
- A coding model explains a patch and another model challenges its assumptions.
- A general model answers a research question and a specialist checks for missing evidence.
- A lower-cost model does most of the work while an expensive model is used only for review.
You can also add advisor instructions such as:
{
"advisor": {
"model": "anthropic/claude-opus-4.8",
"mode": "required",
"instructions": "Focus on data-loss risks and missing rollback steps."
}
}
NanoGPT does not automatically send the advisor your entire conversation. The main model writes a focused question and includes the details it believes the advisor needs.
The tradeoff: more cost and more time
If auto mode decides not to consult, you pay for the normal main-model call.
When a consultation happens, there can be three billable calls: the main model's first pass, the advisor's review, and the main model's final answer. The request will usually take longer too.
That is why Advisor should be used for work where the extra review has a clear purpose, rather than as a switch to turn on for every prompt.
Current availability
Advisor currently works with pay-as-you-go API keys on non-streaming Chat Completions and Responses requests. Some combinations, including subscriptions, BYOK, Private Mode, client tools, and structured outputs, are not supported yet.
The full documentation covers the request options, response details, billing breakdown, and current compatibility rules:
https://docs.nano-gpt.com/api-reference/miscellaneous/advisor
The short version is simple: choose a main model, choose a different reviewer, and give the reviewer a specific question worth answering.
