AWS Bedrock

Usage

import boto3
from tokenid import JournaledBedrock

bedrock = JournaledBedrock(
    boto3.client("bedrock-runtime", region_name="us-east-1"),
    session_id="my-session",
)

resp = bedrock.invoke_model(
    modelId="anthropic.claude-sonnet-4-20250514-v1:0",
    body=b'{"messages":[{"role":"user","content":"Hello"}],"max_tokens":256,"anthropic_version":"bedrock-2023-05-31"}',
)

JournaledBedrock wraps bedrock-runtime and forwards every call. Token counts and cost are journaled to TokenID after the response returns. Works for both Anthropic Claude on Bedrock and Amazon Nova models.

Streaming

events = bedrock.invoke_model_with_response_stream(
    modelId="anthropic.claude-sonnet-4-20250514-v1:0",
    body=b'...',
)
for ev in events["body"]:
    ...

Token counts are aggregated across the final stream chunk and journaled on completion.

Cost basis

Bedrock pricing is parsed from the response metadata (inputTokenCount / outputTokenCount). For Nova models the parser also falls back to the body-level usage.inputTokens / usage.outputTokens.

Pulling historical usage

To backfill historical Bedrock usage from CloudWatch Model Invocation Logs, configure an AWS connection in the TokenID dashboard with an IAM key that has logs:StartQuery and logs:GetQueryResults on the /aws/bedrock/modelinvocations log group. The puller queries Logs Insights every hour and normalises every invocation into the events ledger.