OpenRouter explains three methods for reliable structured LLM output
Overview
OpenRouter's blog explains three ways to get structured output from an LLM: a strict JSON Schema, a forced tool call, or a decision model, with the right choice depending on what the model supports.
In a test of 24 messy support messages run on nine structured-output models plus GLM 5.3 FlashX, 209 of 216 replies matched a strict schema, while prompt-only JSON requests failed 18 times because the replies included code fences.
The method applies only when the request sets response_format to json_schema or json_object, and only on requests that are not streamed. It does not cover tool call arguments, and it cannot fix a reply that max_tokens cut off.
Written by AI from the articles below · updated Nov 10, 12:03 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- OpenRouter BlogOfficialOpenRouter explains three ways to get reliable structured output from any LLM
AIOpenRouter says LLM structured output can be obtained with a strict JSON Schema, a forced tool call, or a decision model, depending on the model's support. In a test of 24 messy support messages on nine structured-output models plus GLM 5.3 FlashX, 209 of 216 replies matched a strict schema, while prompt-only JSON requests failed 18 times because of code fences.
Heat Trend
Not enough continuous observations to show a trend yet.