AWS Bedrock Cost Optimization Guide: Explained

AWS Bedrock is the most expensive idle resource type in your AWS account. A single provisioned throughput unit on Claude or Llama costs roughly $21/hr, that is $15,000/mo per unit running 24/7. PT units are bought for peak load and frequently sit idle on weekends or after a campaign ends. One forgotten PT unit erases a quarter of optimization work elsewhere.

Provisioned throughput is just one of ten Bedrock cost surfaces. Custom models accumulate without invocations. Imported models sit unused. Knowledge bases never get queried. Agents are created and forgotten. Guardrails are configured but never attached. Each is a separate audit rule in ZopNight.

This guide walks through the full Bedrock cost surface and the 10 audit rules ZopNight ships (RC-1601 to RC-1610). It covers PT idle, PT underutilized, PT break-even vs on-demand, custom and imported model orphan and idle detection, KB and agent idle, guardrail unused, invocation logging missing, and deprecated model generation.

This guide keeps the theory short and spends most of its length on what you can actually do. Every recommendation here is one ZopNight can help you execute, starting from a read-only connection.

The big-ticket items: provisioned throughput

Three PT rules cover the highest-value scenarios. RC-1601 detects PT units with zero or near-zero invocations over the analysis window, classic idle. RC-1602 detects PT units with invocations below their PT capacity, underutilized. RC-1603 runs a break-even calculator: compares PT cost against the on-demand token spend that would replace it. If on-demand is cheaper, ZopNight recommends releasing the PT unit. The break-even calculation accounts for input and output token rates per model.

Custom and imported models

RC-1604 detects orphan custom models: models with no PT and no invocations. RC-1605 detects custom models with PT but no traffic, the cost is in the PT unit, not the model itself. RC-1606 detects imported models that have not been invoked in the analysis window. Custom and imported models are cheap to keep around (small storage cost) but indicate dead workflows that often have associated PT or KB resources still incurring cost.

Knowledge bases, agents, guardrails

RC-1607 detects knowledge bases with no retrieval calls, idle. RC-1608 detects agents with no invocations. RC-1609 detects guardrails created but never attached to an agent or model. Each rule fires off invocation count and last-used timestamp. RC-1610 catches deprecated foundation model generations, callers still on prior model versions when newer (cheaper or more capable) versions are available.

Invocation logging

RC-1611 (paired with the rest of the family) flags accounts with Bedrock invocation logging disabled. Without invocation logs, root cause analysis on Bedrock anomalies is blind. ZopNight flags this as a compliance and observability gap, not a direct cost issue, but it is required to run the other Bedrock recommendations effectively.

Key takeaways

• A single Bedrock PT unit costs roughly $21/hr, about $15,000/mo if it runs 24/7.
• 10 Bedrock audit rules (RC-1601 to RC-1610) cover PT idle, underutilized, and break-even, plus orphan models, idle KBs and agents, unused guardrails, and deprecated models.
• The break-even calculator compares PT cost to equivalent on-demand spend so you only keep PT where it is genuinely cheaper.
• Releasing a PT unit is billing-only, agents and applications fall back to on-demand pricing without functional change.

Where ZopNight fits

ZopNight turns this from reading into doing. It ships 490 built-in audit rules across AWS (216), GCP (127), and Azure (147), 124 of those recommendations are wired to act end to end, 28 one-click and 96 guided, and it starts read-only so you can see the opportunity before you act on any of it. The most direct place to begin is scheduling non-production resources to your working hours, which is covered in the FinOps guide and shown concretely for AWS EC2.

How ZopNight schedules non-production resources

The loop that does this is deliberately mechanical, and it starts read-only. You connect your cloud provider with a read-only role, and ZopNight discovers every non-production resources across your regions and accounts. It records a per-action permission verdict for each one, so you can see where it can list a resource but not yet stop it, and you review that inventory, filter it by status or type, and search for the specific resources you care about before anything is scheduled.

Scheduling itself is a cron you write once in plain terms, stop at 7 PM, start at 8 AM on weekdays, pinned to your timezone so the jobs fire at local business hours rather than UTC. A weekly 24-hour grid shows the schedule visually so you catch gaps and overlaps before you save, and an estimate of active versus inactive hours appears before you commit. Resources attach individually or bundle into groups like “dev-cluster” or “staging-db” so a whole environment follows one cadence.

Actions run in dependency order, so a database comes up before the app server that depends on it. When something needs to stay up, an override forces a non-production resources ON or OFF for a defined window, carries a reason so teammates understand why it exists, and expires automatically so nothing is left running by accident. If a start or stop fails, ZopNight retries up to three times and falls back to a dead-letter queue rather than silently dropping the action, and every state change lands in an audit trail that records whether a schedule, an override, or a specific user triggered it.

Beyond the schedule: recommendations, rightsizing, and showback

Scheduling is the fastest lever, but it is one of several. ZopNight ships 490 built-in audit rules across AWS (216), GCP (127), and Azure (147) that flag idle, oversized, and orphaned resources, and each recommendation shows the current monthly cost next to the estimated optimized cost so you act on the largest first. 124 of those recommendations are wired to act end to end, 28 one-click and 96 guided: one-click actions run immediately behind an admin-approval gate, and guided actions add a type-to-confirm review so you check the change before it lands. You mark a recommendation applied once you act, or dismiss the ones that do not fit.

Idle detection reads CPU, network, and connection metrics over a rolling window to separate a genuinely idle resource from one with real but intermittent traffic. Rightsizing is guided and computed from measured utilization over a real window, never a flat 24/7 assumption, so the projected figure matches the bill you actually see. For steady-state fleets, VM autoscaling runs in one of three modes derived from the credential’s permissions: monitor, recommend, or autopilot.

What is left after optimization gets attributed rather than hidden. Showback splits shared cost across owning teams and rolls up by cloud tag, GCP label, or Azure tag, with a Sankey cost-flow view that traces spend across provider, account, type, and team and a savings overlay that points straight at the reclaimable flows. A daily anomaly job writes root-cause markers onto the cost trend, an instance resize, a new resource, a reservation expiry, a failed schedule, so a spike explains itself instead of prompting a manual hunt. And 43 read-only tools expose the same data to an AI assistant over MCP, so you can ask an assistant in Claude, Cursor, or Codex for the same numbers.

Best practices that keep the savings

A few habits separate teams that hold onto the savings from teams that watch them drift back:

• Start with non-production and prove it there. Development, staging, QA, and demo environments carry almost no risk and the largest idle share, so they are the right place to build confidence before anyone considers production.
• Schedule by group, not by hand. Bundling an environment into a group like “staging” means one cadence covers every resource in it, and resources you add later inherit the schedule instead of being quietly forgotten.
• Use overrides instead of disabling schedules. When a late deploy needs a box overnight, a time-boxed override with a written reason keeps the schedule intact and expires on its own, so a one-off exception never becomes a permanent leak.
• Watch the audit trail and notifications. Every start, stop, and failure is logged and can post to Slack, Teams, or Google Chat, so a failed action is visible the moment it happens rather than discovered on the next invoice.
• Treat it as an operating rhythm, not a cleanup. The teams that keep the bill down review recommendations on a cadence and let the automation run continuously, instead of a one-off spring-clean that snaps back the moment attention moves on.

Getting started

Getting started is intentionally low-stakes:

• Connect your cloud provider with a read-only role. Nothing is scheduled or changed at this stage.
• Let ZopNight discover your non-production resources and review exactly what it found, filtered by account, region, and status.
• Create a schedule in your timezone and attach the non-production resources or groups you want it to cover.
• Watch the first cycle run, with Slack, Teams, or Google Chat notifications on every start, stop, and failure, then layer in idle cleanup and guided rightsizing.

Production stays excluded by default throughout, and because discovery and recommendations are read-only, you can prove the value before you enable a single action.