Integration guides ยท 2026-09-22
Building a video analysis flow with video input and reasoning_effort on Step 3.7 Flash
How to combine Step 3.7 Flash's video input support with the low, medium, and high reasoning_effort levels to balance latency and depth of detail in a video analysis flow.
Video input and thinking control on the same model
Step 3.7 Flash's catalog page lists image and video input, function calling, and low/medium/high reasoning_effort levels together; this combination isn't common among models that accept video input. Its 256K context window also allows carrying content from multiple video frames or a long video transcript in a single request.
reasoning_effort levels here come into play not just on text tasks but also when interpreting video input, so they can be used to balance latency and detail in video analysis too.
Which level fits which video task
For detecting a single, clear event in a video (asking whether a specific object appears in a scene, for example), the `low` level is usually enough and responds with low latency. For a more complex analysis like relating multiple events across a video in chronological order, interpreting why a scene happens, or resolving details that seem contradictory, on the other hand, the `medium` or `high` level lets the model evaluate the video content more thoroughly.
Combining a long video input with high reasoning_effort can noticeably extend total response time; in a real-time use case, you need to test that balance with your own video length and task complexity.
- Detecting a single, clear event: the low level is usually enough and fast.
- Relating multiple events, cause-effect interpretation: the medium or high level gives more thorough evaluation.
- Test the combination of long video plus high reasoning_effort in your real-time scenario.
Verify video format and duration limits
Details like supported video formats, maximum duration, or file size can change on the provider side; verify these limits against the model page or current API documentation before going to production.
Frequently asked questions
Does Step 3.7 Flash accept only video, or images too?
It accepts both; the catalog page lists image and video input as separate capabilities. Which one you use in a given request depends on the task.
Can json_mode also be used for video analysis tasks?
The model supports json_mode; if you want to extract a structured summary from video content, you can try it alongside reasoning_effort, but verify the output shape with your own task examples.