What happened?

The cost_per_token() function sometimes returns negative input cost when using cached prompts with bedrock anthropic models. It seems cache_creation_input_tokens and cache_read_input_tokens are subtracted from the prompt_tokens, resulting in a negative value.

Update: This issue seems to be related to claude 4.5. Everything looks legit after downgrading to claude 4.

Relevant log output

# minimal reproduction code
# token counts are taken from the usage object

from litellm import cost_per_token

print(cost_per_token(
  model="us.anthropic.claude-sonnet-4-5-20250929-v1:0", 
  completion_tokens=497,
  prompt_tokens=22475,
  cache_creation_input_tokens=7183,
  cache_read_input_tokens=22465,
))

# Output: (-0.132944625, 0.008200500000000001)

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

v1.77.7

Twitter / LinkedIn details

No response