Part 04

Cost Planning

Pricing and model availability change. Build a workload-specific estimate from current official rates, measured session usage, integration charges, and the human work needed to review and operate the system.

Reviewed 31 Aug 2026 · live rates required · no fixed ROI promise

Two Dimensions of Cost

1. Token Costs

Managed sessions consume model tokens. Model, cache behavior, context size, tool results, retries, and inference settings affect the bill. Read the current pricing page at estimation time.

measure

Use the authoritative usage and list-cost fields emitted for each session.

2. Session Runtime
cap

Session runtime and any separately metered server tools contribute to cost under the current contract. Set a session max_list_cost budget and application limits before sending work.

Source: Anthropic's current pricing and session budget documentation. Contract terms, negotiated rates, taxes, cloud platform charges, and upstream APIs can differ.

Select a Model by Evaluation

Do not preserve a permanent “best model” ranking in course content. Shortlist currently supported models, then choose the least-cost option that clears your task's quality, safety, latency, and reliability gates.

Evaluation sliceMeasureRelease gateEvidence
Representative Accuracy, groundedness, reviewer effort Your task-specific minimums Versioned fixture set and human labels
Adversarial Prompt injection, secret handling, unsafe tool attempts No unauthorized action or data crossing Attack cases plus approval-event audit
Regression Quality, p95 latency, list cost, error rate No material degradation from approved baseline Pinned agent version and reproducible run log

Three Cost Controls

A list-price estimate is a starting point, not a budget or a business case.

Current Rates
LINK
Snapshot the date, model, features, and contract used
Session Budget
HARD CAP
Set at creation; handle budget_reached safely
App Limits
GUARD
Cap elapsed time, turns, tool calls, retries, and request rate

3 Pilot Worksheets

These are planning templates, not measured customer results. Populate them with synthetic test runs, then repeat on an approved, minimized data sample.

Pilot 1: Read-only Ad Report
Synthetic fixtures | current candidate models | no platform writes
  • Inputsaccounts x campaigns x date range
  • Measurelist cost + active time + reviewer minutes
  • Quality gateschema checks + cited anomalies
  • Safety gateread-only scopes + tenant isolation
Decision pilot data required
Forecast = measured p50/p95 cost x approved run volume
Pilot 2: Campaign Draft
Synthetic brief | draft artifact only | human approval required
  • Inputsbrief size + policies + approved sources
  • Measurelist cost + latency + revision time
  • Quality gatepolicy, math, and claim checks
  • Safety gateno spend, publish, or audience write tool
Decision pilot data required
Add human review and integration cost to every forecast
Pilot 3: Event-driven Monitoring
Alert-only | bounded schedule | canary account
  • Inputsevents/day + context/event + lookback
  • Measurelist cost + precision + recall + alert load
  • Quality gatefalse-positive budget + reviewed alerts
  • Safety gatealert only + stop switch + rate cap
Decision pilot data required
Compare scheduled polling with event-triggered execution

Total Cost of Ownership

Do not compare an API subtotal with a person's salary. Compare equivalent, quality-controlled operating models and include integration, review, failure, security, and compliance costs.

Custom Agent Loop
MEASURE
engineering, sandbox, orchestration, on-call, model and tool usage
Self-hosted Managed Agent
MEASURE
worker fleet, hardening, isolation, egress, secrets, retention, incident response
Managed Cloud
MEASURE
platform usage, integrations, evaluation, approvals, monitoring, legal and support

Measure Caching; Do Not Promise Savings

Prompt caching can change input cost, but the realized effect depends on cache-eligible content and hit rate. Compare equivalent runs and read the current pricing rules.

Prompt Caching Impact
Baseline record usage + quality
Candidate same fixtures + cache telemetry
Decision measured delta, not assumed

Keep model and Agent version, fixture set, tool policy, and inference settings constant when comparing. A lower bill is not a win if quality, safety, or reviewer effort regresses.