AWS Bedrock cost playbook: attribute, log, cap, route, review

On 30 October 2026, AWS published a playbook for FinOps and platform teams on managing Amazon Bedrock costs, written by AWS specialist solutions architects and product staff. It argues that the fix for rising generative AI bills is neither a cheaper model nor a shorter prompt but an operating model: attribute every dollar to an owner, log every request, enforce budgets, set efficient defaults, and review on a cadence. It is vendor guidance with no customer results, and it publishes no cost figures from a real deployment.

Why agents change the bill

The authors say a chat assistant makes one model call per user turn, while an agent such as a coding assistant plans, reads files, calls tools and iterates, making many calls per task with a larger context each time. Spend then follows how much autonomy a workflow has and how much context it carries, not how many people use it. They say published averages are a poor guide and teams should measure their own consumption.

Stage 1: the shared-role problem

AWS says Bedrock records the calling IAM identity in the line_item_iam_principal column of the Cost and Usage Report 2.0. The catch, according to the post, is a gateway that reaches Bedrock through a single role: every dollar lands on that role and per-user visibility disappears. The remedy is for the gateway to assume the Bedrock role once per user, pass the user as the role session name and their team, application and cost centre as session tags, and cache those credentials for up to 1 hour. AWS adds that billing attribution lags roughly 24 hours, so near-real-time dashboards need invocation logs and OpenTelemetry.

Stage 3: three ways to enforce a budget

• Tier 1, monitor and alert: AWS Budgets alerts on the billing data. Nothing to deploy and no added latency, but nothing can hard-stop spend. AWS says it suits early adoption.
• Tier 2, IAM deny on breach: a Lambda function tallies per-user tokens and dollars in DynamoDB from invocation logs and adds a breaching user to a shared deny policy. AWS says a block lands seconds to minutes after the breaching call, and that checks based on client telemetry can differ from the billing report.
• Tier 3, gateway: a component in the inference path returns HTTP 429 the moment a limit is crossed, with per-user model allowlists. The trade-off AWS names is a component that must be deployed, scaled and security-reviewed, with clients configured to use it.

The authors say that if developers call Bedrock directly, only Tiers 1 and 2 are available, and that most organisations should start at Tier 1. They say that if spend is already in the millions per month, build enforcement now.

Stage 4: routing and caching

On caching, the post cites Bedrock's published US East (N. Virginia) pricing as of October 2026: a cache write costs 1.25 times the base input rate with a 5-minute lifetime (2 times with 1 hour); a read costs 0.1 times on Claude Sonnet 4.6 and Haiku 4.5, and 0.05 times on Claude Sonnet 5.5 and Opus 5.5. AWS's arithmetic is that one reuse pays for the write, and ten calls on a stable prefix cost about 22% of the uncached price at the 0.1 rate and 17% at 0.05. A low read ratio usually means something like a timestamp sits before the cache point. For routing, it says more than half of calls on cheaper models is an achievable target for coding fleets, after a quality evaluation on the top workflows.

What the post leaves out

There are no measured savings, no cost for running the Lambda or gateway tiers, and no data on how often the thresholds trigger. The conclusion concedes that judging whether a workflow is worth running means comparing the agent with the human process it replaced, which the playbook does not cover.