Amazon Bedrock and Anthropic report different usages

[Pranav-Wadhwa]

Description

I've noticed that the logic for calculating inputTokens and inputTokenDetails is inconsistent for claude between Anthropic as the provider and Amazon Bedrock as the provider. Here's an example script:

import "dotenv/config"
import logger from "@/utils/logger"
import { generateText } from "ai"
import { anthropic } from "@ai-sdk/anthropic"
import { createAmazonBedrock } from "@ai-sdk/amazon-bedrock"
import { env } from "@/env.mjs"

const bedrock = createAmazonBedrock({
  region: "us-east-2",
  accessKeyId: env.accessKeyId,
  secretAccessKey: env.secretAccessKey,
})

// System prompt must be at least 1024 tokens for Anthropic caching to work
const systemPrompt = `<a long system prompt...>`

const userMessage = "What is 2 + 2? Reply with just the number."

async function testAnthropicProvider() {
  logger.info("=== Testing Anthropic Provider ===")

  const result = await generateText({
    model: anthropic("claude-sonnet-4-5-20250929"),
    messages: [
      {
        role: "system",
        content: systemPrompt,
        providerOptions: {
          anthropic: { cacheControl: { type: "ephemeral" } },
        },
      },
      {
        role: "user",
        content: userMessage,
      },
    ],
  })

  logger.info("Anthropic response", { text: result.text })
  logger.info("Anthropic usage", { usage: result.usage })
}

async function testBedrockProvider() {
  logger.info("=== Testing Bedrock Provider ===")

  const result = await generateText({
    model: bedrock("global.anthropic.claude-sonnet-4-5-20250929-v1:0"),
    messages: [
      {
        role: "system",
        content: systemPrompt,
        providerOptions: {
          bedrock: { cachePoint: { type: "default" } },
        },
      },
      {
        role: "user",
        content: userMessage,
      },
    ],
  })

  logger.info("Bedrock response", { text: result.text })
  logger.info("Bedrock usage", { usage: result.usage })
}

export default async function main() {
  try {
    await testAnthropicProvider()
  } catch (error) {
    logger.error("Anthropic error", { error })
  }

  try {
    await testBedrockProvider()
  } catch (error) {
    logger.error("Bedrock error", { error })
  }
}

logger.info(`${scriptName}...`)
await main()
logger.info(`${scriptName}... done`)

Here is the Anthropic usage:

{
  "inputTokens": 2590,
  "inputTokenDetails": {
    "noCacheTokens": 21,
    "cacheReadTokens": 0,
    "cacheWriteTokens": 2569
  },
  "outputTokens": 5,
  "outputTokenDetails": {},
  "totalTokens": 2595,
  "raw": {
    "input_tokens": 21,
    "output_tokens": 5,
    "cache_creation_input_tokens": 2569,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 2569,
      "ephemeral_1h_input_tokens": 0
    },
    "service_tier": "standard"
  },
  "cachedInputTokens": 0
}

Here is the Bedrock usage

{
  "inputTokens": 21,
  "inputTokenDetails": {
    "noCacheTokens": 21,
    "cacheReadTokens": 0,
    "cacheWriteTokens": 2569
  },
  "outputTokens": 5,
  "outputTokenDetails": {
    "textTokens": 5
  },
  "totalTokens": 26,
  "raw": {
    "inputTokens": 21,
    "outputTokens": 5,
    "totalTokens": 2595,
    "cacheReadInputTokens": 0,
    "cacheWriteInputTokens": 2569
  },
  "cachedInputTokens": 0
}

Anthropic includes the cacheReadTokens and cacheWriteTokens in the inputTokens whereas Bedrock does not. Additionally Bedrock computes the noCacheTokens as inputTokens - cacheReadTokens which often times results in a negative value.

AI SDK Version

"@ai-sdk/amazon-bedrock": "^4.0.4",
"@ai-sdk/anthropic": "^3.0.1",
"ai": "^6.0.3",

Code of Conduct

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