Skip to main content
Available in arize-phoenix 15.8.0+ The Phoenix Playground now surfaces dedicated controls for extended thinking on Anthropic and Google models. Thinking parameters are persisted with saved prompts and restored correctly on reload or provider switch.

​
Anthropic Extended Thinking

Three controls appear when an Anthropic model is selected:
  • Thinking — choose between Adaptive (model decides when to think), Enabled (always think), or disabled. Adaptive mode is the default for new instances.
  • Thinking Budget — token budget allocated for the thinking block. Minimum is 1,024 tokens; maximum is capped by the max_tokens setting. Defaults to 5,000.
  • Thinking Display — toggle whether the thinking block is shown inline in the Playground response panel.
Anthropic also exposes an Effort control (low / medium / high) for models that support output-confidence effort levels independently of extended thinking.

​
Google Thinking

For Google Gemini models:
  • Thinking Budget — maximum tokens the model may spend on internal reasoning before emitting the response.
  • Thinking Level — preset effort level (LOW, MEDIUM, HIGH). Overrides the budget when set.
Thinking parameters are saved as part of the prompt version and round-tripped through storage and the SDK export format. Switching to a provider that does not support thinking drops the thinking config automatically; switching back restores it.