Not streaming tool input when using anthropic provider

[cloudycotton]

Description

I noticed when using anthropic provider, the stream is not smooth and updating in very big chunks compared to openai or any other provider available.

AI SDK Version

latest version of AI SDK.

Code of Conduct

• I agree to follow this project's Code of Conduct

[gr2m]

Can you verify if this is an AI SDK or an Anthropic API problem?

[BeiXiao]

same issues

[cloudycotton]

This is an AI sdk problem. Just have a simple write_file tool and call it using anthropic via gateway or native provider and stream on the frontend. It won't stream the file writing input updates like openai. If you downgrade to sdk v5 beta, it does.

[cloudycotton]

import {  streamText, tool } from "ai"
import { createAnthropic } from "@ai-sdk/anthropic"
import z from "zod"

const anthropic = createAnthropic({
    apiKey: process.env.ANTHROPIC_API_KEY,
})

const response =  streamText({
    model: anthropic("claude-sonnet-4-20250514"),
    tools: {
        write_file: tool({
            name: "write_file",
            description: "Write a file to the file system.",
            inputSchema: z.object({
                path: z.string().describe("The path to write the file to."),
                content: z.string().describe("The content to write to the file.")
            }),
            execute: async ({ path, content }) => {
                return {
                    content: "File written successfully."
                }
            }
        })
    },
    messages: [
        {
            role: "user",
            content: "Write a 5000 line long text in 3 html files."
        },
    ]
})

@gr2m

Here is an example. When you stream the input of write_file on the frontend, with anthropic. It streams in very big chunks or not at all whereas with gpt provider, it streams it in a smooth manner. This same code worked differently in beta version of ai sdk.

[gr2m]

Thanks for providing that test case, I was able to reproduce the problem, the execute callback was not called at all for me with the anthropic provider. Could you check if it works correctly using @anthropic-ai/sdk? That would be my next step, to make sure it's not a problem outside of the AI SDK. It's possible that Anthropic made a change, so that it worked a while ago with theAI SDK while in beta, but then stopped working suddenly.

[BeiXiao]

Thanks for providing that test case, I was able to reproduce the problem, the execute callback was not called at all for me with the anthropic provider. Could you check if it works correctly using @anthropic-ai/sdk? That would be my next step, to make sure it's not a problem outside of the AI SDK. It's possible that Anthropic made a change, so that it worked a while ago with theAI SDK while in beta, but then stopped working suddenly.

Is it related to the “Fine-grained tool streaming” setting? https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/fine-grained-tool-streaming

[gr2m]

Interesting. Looks like we do not permit setting the betas field explicitly, but we do use it internally. We might want to add betas to provider options for Anthropic, or introduce an explicit option like providerOptions.anthropic.experimental_fineGrainedToolStreaming.

[gr2m]

with anthropic. It streams in very big chunks or not at

I couldn't get it to stream or even call the execute() callback at all.

If you downgrade to sdk v5 beta, it does

I tried to downgrade to ai@5.0.0-beta.34 and it still didn't work. I tried ai@5.0.0-beta.20 as well but to no avail.

[cloudycotton]

The version it worked for us with native anthropic provider was 5.0.0-beta.3

…

On Mon, 8 Sep 2025 at 10:21, Gregor Martynus ***@***.***> wrote: *gr2m* left a comment (vercel/ai#8453) <#8453 (comment)> with anthropic. It streams in very big chunks or not at I couldn't get it to stream or even call the execute() callback at all. If you downgrade to sdk v5 beta, it does I tried to downgrade to ***@***.*** and it still didn't work. I tried ***@***.*** as well but to no avail. — Reply to this email directly, view it on GitHub <#8453 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BIJLQNM5NLIUK5P3JOIK2A33RUDMNAVCNFSM6AAAAACFVMLKTWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENRUGU4TKNJRHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[gr2m]

are you sure the execute function is called for you? I tried it with beta.3 with the code below, but to no avail. I see an inital text streamed like this

I'll create 3 HTML files with approximately 1,667 lines each to reach the 5,000 line total. Each file will contain structured HTML content with meaningful text.

then 30s+ nothing, and then it just ends.

import { anthropic } from '@ai-sdk/anthropic';
import { streamText, tool } from 'ai';
import z from 'zod';
import 'dotenv/config';

async function main() {
  const result = streamText({
    model: anthropic('claude-sonnet-4-20250514'),
    tools: {
      write_file: tool({
        name: 'write_file',
        description: 'Write a file to the file system.',
        inputSchema: z.object({
          path: z.string().describe('The path to write the file to.'),
          content: z.string().describe('The content to write to the file.'),
        }),
        execute: async ({ path, content }) => {
          console.log('EXECUTE!')
          return {
            content: 'File written successfully.',
          };
        },
      }),
    },
    messages: [
      {
        role: 'user',
        content: 'Write a 5000 line long text in 3 html files. Use the provided "write_file" tool to write the files.',
      },
    ],
  });

  for await (const textPart of result.textStream) {
    process.stdout.write(textPart);
  }

  console.log();
  console.log('Token usage:', await result.usage);
  console.log('Finish reason:', await result.finishReason);
}

main().catch(console.error);