I noticed when using gpt-4-vision-preview and sending image parts, the token calculations are negative. The token counts are positive when sending messages with no image parts, and using other models like Anthropic seem to count tokens correctly.
oh this means Helicone ran into an error getting the tokens. Would you mind if I took at look at your org? Can you DM me your org name and add me to your org? justin@helicone.ai?