Models JSON Schema Support

[rajveerappan]

Select Topic Area

Question

Body

Models added json schema response format support back in January for just 4o and that's still the case even though 4o-mini and all the 4.1 variants including the mini/nano ones support json schema via the OpenAI API.

What is the reason the schema support is limited to 4o and is there a plan to expand that support?

[Harshit-Maurya838]

Hey, great question!

Yeah, it’s true that as of now, JSON schema response format is only officially supported in GPT-4o via the OpenAI API. Even though models like 4o-mini and others in the 4.1 family exist, they currently don’t expose that schema support publicly.
From what I understand, it's likely a combination of performance optimization and rollout control. 4o was designed with JSON mode and tool use as a core part of its functionality, whereas the others might not yet have the same level of stability or internal support for strict schema handling.
That said, I haven’t seen any official announcement about expanding it to the mini/nano models, but considering the demand, I’d imagine it's at least on the roadmap. Would be awesome to have that feature on lighter models too.
Fingers crossed they open it up soon.

[rajveerappan]

Structured outputs are definitely supported in all models since 4o, including 4o-mini and all the 4.1 family: https://platform.openai.com/docs/guides/structured-outputs?api-mode=chat#supported-models

Structured Outputs is available in our latest large language models, starting with GPT-4o.

[KateCatlin]

Hey @rajveerappan, thanks for nudging us on this! I'll spin up an issue to get this supported on all applicable models.

Can I ask - are you building something with 4o-mini or a 4.1 variant that requires specific JSON schema support? Always curious what our users are building :)

[rajveerappan]

thank you @KateCatlin! if it's a public issue could you please tag me on it so i can follow along? 🙏

as far as use case we've found that in essentially every use case where the LLM output needs to be consumed by code structured outputs is the only way to go. however different models perform at different levels based on the complexity of the JSON Schema so the models product would be a great way to compare their performance side by side.

similarly it would be great if the evals specifically worked for comparing JSON output. for example you could have one model output key a before b and the other could output key b before a but as long as the values of both keys in the JSON match the eval should pass since order is not important when comparing JSON

[rajveerappan]

@KateCatlin are there any public issues i can follow along for this?

[nakulgadhewal050]

OpenAI added JSON schema response format support in January 2024 for the gpt-4o model in the Chat Completions API, but as of now, that specific support is not available for the gpt-4.0-turbo, gpt-4.0-turbo-mini, gpt-4.1, gpt-4.1-mini, or gpt-4.1-nano models via the same interface.

[garv999]

Select Topic Area

Question

Body

Models added json schema response format support back in January for just 4o and that's still the case even though 4o-mini and all the 4.1 variants including the mini/nano ones support json schema via the OpenAI API.

What is the reason the schema support is limited to 4o and is there a plan to expand that support?

JSON schema support is limited to GPT-4o because it was specifically optimized for reliable schema outputs. Other models like 4o-mini and 4.1 variants weren’t initially fine-tuned for strict schema adherence. The feature is rolling out gradually—it's already supported via API for newer models, and wider product support will likely come in future updates.