Assistant: Is Anthropic cache working correctly?

[andreatitolo]

Hi!
Really liking the assistant so far! One thing I noticed is that while the Anthropic cache system is working (as per multiple merged PR), I still see a large use of input tokens instead of cache read/write in Anthropic console when using the Claude through the chat, mostly when dealing with long conversations.

You can see a comparison in these two pictures below.

This is one using the Continue extension, and a long conversation of around 3,030,010 tokens

Expand

This is instead a conversation with positron assistant, little more that half the tokens of the continue one (1,764,200), but the less use of cache results in almost the same cost as the previous one.

Expand

I know comparing tools is not exactly the best as each has its own implementation, but that seems quite a strong difference.

I can open an issue if appropriate.

Positron version and OS

Positron Version: 2025.07.0 build 204
Code - OSS Version: 1.100.3
Commit: https://github.com/posit-dev/positron/commit/03ae7a5393c776bb24c23d2aa6a6bfbba90cbc5e
Date: 2025-06-30T16:29:03.492Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.5.0

[juliasilge]

Most of those caching PRs went in after we set up the last regular monthly release. Would you be up for switching to a daily build, to see if you still observe that same difference?

If that's not a good fit for you, our next regular monthly release will go out around Aug 4.

[andreatitolo]

I see, thanks for pointing that out :) I will switch to a daily build and check in the next days :)