What happened?
Anthropic API updated their spec at some point to include full usage details in messageDelta events, previously it used to be just output tokens. Now with the update, it includes input and cache token counts. These should be counted cumulatively with tokens from messageStart events as quoted on the referenced page:
The token counts shown in the usage field of the message_delta event are cumulative.
This results in some APIs following Anthropic's API spec breaking Cline's token and context window tracking, resulting in a degraded user experience. The fix for this would require upgrading the Anthropic SDK as it needs the new MessageDeltaUsage definitions that includes the new count types.
Reference: https://docs.anthropic.com/en/docs/build-with-claude/streaming
Steps to reproduce
Using Anthropic API, token counts may be off. This may or may not happen on the official API spec depending upon how they count and communicate tokens. I am able to reproduce it in my local environment through a custom proxy service that exposes Anthropic compatible API.
Relevant API REQUEST output
Provider/Model
Anthropic Models
Operating System
macOS 15.5
System Info
MacBook Pro: M1 Max 64GB
Cline Version
3.17.14-a
Additional context
I have a PR to fix this issue, creating this issue to track it.
What happened?
Anthropic API updated their spec at some point to include full usage details in
messageDeltaevents, previously it used to be just output tokens. Now with the update, it includes input and cache token counts. These should be counted cumulatively with tokens frommessageStartevents as quoted on the referenced page:This results in some APIs following Anthropic's API spec breaking Cline's token and context window tracking, resulting in a degraded user experience. The fix for this would require upgrading the Anthropic SDK as it needs the new
MessageDeltaUsagedefinitions that includes the new count types.Reference: https://docs.anthropic.com/en/docs/build-with-claude/streaming
Steps to reproduce
Using Anthropic API, token counts may be off. This may or may not happen on the official API spec depending upon how they count and communicate tokens. I am able to reproduce it in my local environment through a custom proxy service that exposes Anthropic compatible API.
Relevant API REQUEST output
Provider/Model
Anthropic Models
Operating System
macOS 15.5
System Info
MacBook Pro: M1 Max 64GB
Cline Version
3.17.14-a
Additional context
I have a PR to fix this issue, creating this issue to track it.