Mastra•3mo ago

anthropic Prompt caching is not working with mastra

my code:

this.agent = new Agent({
  name: 'Agent',
  instructions: [
    {
      role: 'system',
      content: prompt(),
      providerOptions: {
        anthropic: { cacheControl: { type: 'ephemeral' } },
      },
    },
  ],
  model: wrappedModel,
})

this.agent = new Agent({
  name: 'Agent',
  instructions: [
    {
      role: 'system',
      content: prompt(),
      providerOptions: {
        anthropic: { cacheControl: { type: 'ephemeral' } },
      },
    },
  ],
  model: wrappedModel,
})

this.agent = new Agent({
  name: 'Agent',
  instructions: [
    {
      role: 'system',
      content: prompt(),
      providerOptions: {
        anthropic: { cacheControl: { type: 'ephemeral' } },
      },
    },
  ],
  model: wrappedModel,
})

this.agent = new Agent({
  name: 'Agent',
  instructions: [
    {
      role: 'system',
      content: prompt(),
      providerOptions: {
        anthropic: { cacheControl: { type: 'ephemeral' } },
      },
    },
  ],
  model: wrappedModel,
})

I used it the exact way Mastra documented it in the code, but I’m still seeing usage as 0 :

inputTokens: 90,567
outputTokens: 1,485
totalTokens: 92,052
cachedInputTokens: 0

I also tested the AI-SDK with caching , and caching works there, but not in Mastra. Am I doing something wrong? Did it work for anyone else?

メ

メグディップ•10/3/25, 8:26 AM

I am also having the same issue, tried the transform params middleware and still caching isn't working fully

メ

メグディップ•10/3/25, 8:26 AM

import type { LanguageModelV1Middleware } from "ai";

export const anthropicCachingMiddleware = (cacheAll: boolean = true): LanguageModelV1Middleware => ({
    transformParams: async ({ params }: { params: any }) => {
        const prompt = params.prompt;

        for (const message of prompt) {
            if (
                message.role === "system" ||
                (cacheAll && prompt[prompt.length - 1] === message)
            ) {
                message.providerMetadata = {
                    ...message.providerMetadata,
                    anthropic: { cacheControl: { type: "ephemeral" } },
                };
            }
        }

        return params;
    },
});

import type { LanguageModelV1Middleware } from "ai";

export const anthropicCachingMiddleware = (cacheAll: boolean = true): LanguageModelV1Middleware => ({
    transformParams: async ({ params }: { params: any }) => {
        const prompt = params.prompt;

        for (const message of prompt) {
            if (
                message.role === "system" ||
                (cacheAll && prompt[prompt.length - 1] === message)
            ) {
                message.providerMetadata = {
                    ...message.providerMetadata,
                    anthropic: { cacheControl: { type: "ephemeral" } },
                };
            }
        }

        return params;
    },
});

It's a bit outdated, but similar to this middleware

Mastra Triager•10/3/25, 12:04 PM

Created GitHub issue: https://github.com/mastra-ai/mastra/issues/8456

GitHub

[DISCORD:1423553280910426122] anthropic Prompt caching is not worki...

This issue was created from Discord post: https://discord.com/channels/1309558646228779139/1423553280910426122 my code: this.agent = new Agent({ name: 'Agent', instructions: [ { role: '...

Ariel Weinberger•10/15/25, 2:52 AM

@sarah @Aisha @Meghdip Karmakar

Got this to work:

export const anthropicCachingMiddleware = ({
  overrideModelMessages,
}: {
  overrideModelMessages: ModelMessage[];
}): LanguageModelMiddleware => ({
  transformParams: async ({ params }) => {
    return {
      ...params,
      prompt: overrideModelMessages,
    };
  },
});

export const anthropicCachingMiddleware = ({
  overrideModelMessages,
}: {
  overrideModelMessages: ModelMessage[];
}): LanguageModelMiddleware => ({
  transformParams: async ({ params }) => {
    return {
      ...params,
      prompt: overrideModelMessages,
    };
  },
});

import type { LanguageModelV1Middleware } from "ai"; export const anthropicCachingMiddleware = (cacheAll: boolean = true): LanguageModelV1Middleware => ({ transformParams: async ({ params }: { params: any }) => { const prompt = params.prompt; for (const message of prompt) { if ( message.role === "system" || (cacheAll && prompt[prompt.length - 1] === message) ) { message.providerMetadata = { ...message.providerMetadata, anthropic: { cacheControl: { type: "ephemeral" } }, }; } } return params; }, });

export const anthropicCachingMiddleware = ({ overrideModelMessages, }: { overrideModelMessages: ModelMessage[]; }): LanguageModelMiddleware => ({ transformParams: async ({ params }) => { return { ...params, prompt: overrideModelMessages, }; }, });

anthropic Prompt caching is not working with mastra

Similar Threads

Similar Threads

Similar Threads