fix: improve input token handling for anthropic metadata usage

[caozhiyuan]

What does this PR do?

fix #7481

How did you verify your code works?

run bun dev . and input prompt

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[caozhiyuan]

https://github.com/vercel/ai/blob/ai%405.0.97/packages/anthropic/src/anthropic-messages-language-model.ts#L1ai269
ai sdk anthropic set inputTokens and cachedInputTokens at message_start.
If inputTokens is 0, taking from rawUsage at message_stop https://github.com/vercel/ai/blob/ai%405.0.97/packages/anthropic/src/anthropic-messages-language-model.ts#L1329.

[hank9999]

Hi, my fixes to ai-sdk/anthropic has been merged, only need to update it to 3.0.10, it will be fix. I will do update pr asap

[caozhiyuan]

@hank9999 Upgrading from v2 to v3 seems to be a breaking change

[hank9999]

@hank9999 Upgrading from v2 to v3 seems to be a breaking change

maybe, but i have seen v3 sdk in project.

opencode/packages/opencode/package.json

Line 61 in 05019da

"@ai-sdk/google-vertex": "3.0.97",

opencode/packages/opencode/package.json

Line 53 in 05019da

"@ai-sdk/amazon-bedrock": "3.0.73",

I think it's not too hard to do update.

[caozhiyuan]

@hank9999 I noticed that the data structure of ai sdk usage has changed between v2 and v3 versions; I'll need to review the related code when I have time. Or you could take a look and help submit a PR, much appreciated

https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v2/language-model-v2-usage.ts#L7

https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v3/language-model-v3-usage.ts

[hank9999]

@caozhiyuan Hi, I find a sad fact, there were a version number mass in ai-sdk. Actually, all packages in project are V2, although the version number is 3.0, such as ai-sdk/[email protected], it's V2!

[hank9999]

This PR is the best solution in this case. ):

[caozhiyuan]

@rekram1-node can help review and merge it ?