thinking token
Tokens occupied by the internal reasoning chain that a reasoning-capable large model produces before generating its final answer. They are billed as output tokens, are hidden from or collapsed for the user, and the depth of reasoning is controlled through the thinking_level parameter.
Timeline (last 90 days)
The way thinking tokens are counted (whether truncated reasoning is included, whether the tokenizer is unified, whether comparisons are made at the same sampling temperature) affects the measured compression ratio, and under different definitions the compression ratio may differ by tens of percentage points.
Although thinking tokens of some models are not directly shown to users, they are still counted toward consumption.
All Facts (2)
The way thinking tokens are counted (whether truncated reasoning is included, whether the tokenizer is unified, whether comparisons are made at the same sampling temperature) affects the measured compression ratio, and under different definitions the compression ratio may differ by tens of percentage points.
50%UnverifiedAlthough thinking tokens of some models are not directly shown to users, they are still counted toward consumption.
50%