Hi,
We’ve been playing with o3-mini the last few days. In the playground and in our API we can never get o3-mini to produce more than one tool-call response no matter the functions, tool_choice or reasoning. Is it intentional that o3-mini can only return one tool per API call
Thanks for all the help!
Same problem. Makes it hard to migrate from 4o-mini
o3-mini tool use is a known issue and is being worked on, thanks for taking the time to flag.
Yep, even with a “strict”:“false” function, the AI model just is unable to comply no matter how clear the understanding, because:
This is a tool wrapper that low-iq models have a bad history with…
An alternate strategy is to have a function with an array of the API objects you need, of undetermined length:
This is the function to implement this.
{
"name": "roll_dice",
"description": "Rolls true random dice, supporting multiple dice rolls, returning totals of one trial",
"strict": true,
"parameters": {
"type": "object",
"required": [
"trials"
],
"properties": {
"trials": {
"type": "array",
"description": "Array of trials for rolling dice",
"items": {
"type": "object",
"properties": {
"rolls": {
"type": "number",
"description": "The number of dice to roll"
},
"sides": {
"type": "number",
"description": "The number of sides on each die"
}
},
"required": [
"rolls",
"sides"
],
"additionalProperties": false
}
}
},
"additionalProperties": false
}
}
Out of curiosity, how did you find this out Foxalabs?
I have made a similar post with other problems it is having with tool calls (I cant post links even to this same forum 
):
Foxalabs is a moderator here and we chat with him regularly (if you reach trust level 4, you can become a moderator too) .
Just been playing with o4-mini in the playground. I’m sstill struggling with it returning more than one tool. Is this an issue for the new reasoning models?
Ran into this as well today on o4-mini
The ability to call multiple user tools in parallel is either there – or not.
Parallel tool use also has had complications with “strict” structured output functions.
How does it work? A special tool is placed, alongside the “functions” tool: a recipient for processing those parallel function calls.
This is how it looks to the AI on gpt-4.1-mini on the Chat Completions endpoint:
This is the availability of such to the o4-mini API model:
No wrapper for making parallel function calls. Thus multiple iterations to your API backend and input billing for each context use for individual function calls on Chat Completions (but a token savings by not having that additional tool always there).
You can see above I had to fill in two separate function calls to get my “parallel” use needs.
That same prompt with o4-mini on Responses put the internal iterator into a loop, returning a single function call only after 3354 output tokens. Then failure returning the single function’s value - because apparently the Responses Playground is trying to now resend and rebloat context with a prior reasoning message ID in the messages list with tool return also, needing the same state storage as reusing a response ID - prior response ID which it doesn’t do.
If one is anointed from on high, you mean 
Are you handing those out without any obligation to check a moderator inbox?
