Temperature

Text Generation Process

• Claude tokenizes input text (breaks it into smaller chunks)
• Prediction phase: assigns probability to each possible next token
• Sampling phase: selects a token based on these probabilities
• Process repeats until the full response is complete

Temperature Parameter

• Decimal value between 0 and 1 that controls probability distribution
• Passed to the model when making API calls
• Influences which token Claude selects next

Temperature Effects

• Temperature = 0: Deterministic output; always selects highest probability token
• Temperature increases: Lower probability tokens become more likely to be selected
• Creates a spectrum from predictable to creative outputs

When to Use Different Temperatures

Low temperature (closer to 0):

• Data extraction tasks
• Information retrieval
• Tasks requiring accuracy and consistency
• When you want deterministic, reliable output

High temperature (closer to 1):

• Creative writing
• Brainstorming
• Marketing copy
• Joke generation
• Tasks where unexpected word choices are valuable

Implementation

• Add temperature as a parameter to function calls
• Include it in the params object sent to the model
• Default value of 1. 0 leans toward creativity

Important Note

• Increasing temperature increases the chances of different outputs, but doesn't guarantee dramatically different results
• Multiple runs at the same temperature may still produce similar outputs

Takeaways

• Claude generates text by tokenizing input, assigning probabilities to possible next tokens, and sampling based on those probabilities until complete
• Temperature (0–1) controls the probability distribution: 0 = deterministic/highest probability token always selected, higher values = lower probability tokens become more likely
• Use low temperature (near 0) for accuracy-critical tasks like data extraction; use high temperature (near 1) for creative tasks like writing and brainstorming
• Temperature increases chances of varied outputs but doesn't guarantee dramatically different results across multiple runs
• Temperature is passed as a parameter in API calls, with a default value of 1. 0