How many tokens is the best to extract the best output from the model
There is no magic number for tokens, my recommendations, for
Simple Tasks, for example, Text parsing, basic mapping, is 500 to 1000 max tokens, as it is fast and low-cost, for Complex Logic / Code Generation, I would recommend 2000 - 4000 max tokens, as it gives the model enough room to output well-commented JSON, etc. For advance multi-agent loops, leave it to the model, which allows the model to naturally choose its output length up to its native maximum capability. I hope this helps you 
-
Classification or yes/no answers: 50 to 200
-
Extracting structured data (JSON, fields): 300 to 800
-
Summaries or short analysis: 500 to 1500
-
Long-form content or detailed reasoning: 1500 to 4000
If you’re getting cut-off responses, double your current limit and retest. If cost is a concern, log the actual token usage from the model’s response and dial the ceiling down to 20% above your average.
What are you using it for? That’ll help narrow it down.