Price the whole accepted workflow outcome

Learning objective

Model total cost of ownership per workflow: inference, retrieval, tool infrastructure, retries, evaluation, human review, integration, and maintenance.

Core idea

The cheapest model call does not necessarily produce the cheapest business outcome. A workflow pays for failed attempts, retrieval and tools, review, and the systems that make delivery possible. Divide that complete cost by the accepted outcomes defined in 1.6.1, using a consistent accounting period.

Anthropic’s cost guidance recommends comparing models by cost per completed task. It explicitly includes the tokens spent on a failed task, its retry, and downstream failure costs. Extend that comparison into a workflow cost ledger: one record of variable costs, allocated shared costs, and lifecycle costs, with assumptions finance can inspect.

How it works

Establish the accounting boundary

Start with the workflow, period, and accepted outcome. A monthly ledger might cover requests from receipt through final delivery, including human escalation and correction. Preserve a workflow identifier across all attempts so that a retry adds cost to the same case rather than appearing as another success.

Specify whether you are reporting operating cost, allocated total cost, or cash required to launch. They answer different questions. Amortizing a $48,000 integration over twenty-four months contributes $2,000 to a monthly comparison, but the launch still requires the initial funding. State the allocation period and expected useful life; do not hide the commitment behind a small monthly number.

Give each cost the workflow incurs one ledger line:

These are accounting categories, not separate Anthropic invoice fields. A retry belongs to the inference or tool charges it actually incurred; show its subtotal for analysis without adding the same expense twice. Retrieved text can incur retrieval-service cost and input-token cost because those are different resources. A single search invoice must not be counted under both retrieval and tool infrastructure.

Price inference from measured usage

For Claude API requests, use the response usage rather than characters or guessed conversation length. Anthropic’s documented cost calculation prices usage.input_tokens, usage.output_tokens, and usage.cache_read_input_tokens separately. Cache writes are split into usage.cache_creation.ephemeral_5m_input_tokens and usage.cache_creation.ephemeral_1h_input_tokens when those categories are present.

Multiply each token category by its applicable price per million tokens, divide by one million, and sum across all requests in the workflow. Use the pricing applicable to the selected model and request configuration. The pricing table distinguishes base input, output, cache writes, and cache reads; one blended token price loses those distinctions. Do not add the aggregate cache_creation_input_tokens again after pricing its duration-specific components.

Tool definitions and generated tool-use content contribute token charges. Anthropic’s pricing page also distinguishes client-side tools from server-side tools with additional usage charges. Your client-side execution infrastructure still belongs in the ledger even when the Claude API invoice contains only the model’s token charges for that interaction.

Optimization mechanics belong to the cost-control objectives elsewhere in the guide. Here, record the measured mix of cache reads, writes, output, and attempts. A projected cache hit rate is an assumption to stress-test, not a discount to apply uniformly to every input token.

Reconcile without mistaking reports for TCO

For a Claude Console organization, the Usage and Cost Admin API exposes /v1/organizations/usage_report/messages and /v1/organizations/cost_report. Cost reporting supports group_by[]=workspace_id and group_by[]=description, with daily granularity. Its USD amounts are decimal strings in cents; normalize units before joining them with a dollar-denominated ledger.

Use an Admin API credential for reconciliation; workspace-scoped API keys do not authorize these endpoints. Individual accounts cannot use the Admin API. The documented programmatic Usage and Cost endpoints are unavailable on Claude Platform on AWS, where the Console Usage and Cost pages provide the information instead. Claude Enterprise uses a separate Analytics API. The cost endpoint also excludes Priority Tier costs, so a ledger using that billing model needs its separate cost input.

Allocate shared spend deliberately. If several workflows share a workspace, the daily workspace total alone cannot attribute cost to each one. Join request-level usage estimates with workflow records and explain any reconciliation difference. Reports establish vendor spend; review time, integration, and successful delivery come from your own operational and financial records.

Key distinctions and trade-offs

Cost per attempt versus accepted outcome: An attempt consumes resources regardless of whether it succeeds. The denominator should reflect the result the business buys.

Variable versus committed cost: A model change can lower variable inference cost without changing an integration commitment or review staffing. Report both so decision-makers know which savings are realizable.

Mean versus costly tail: A few long, repeatedly corrected cases can dominate spend. Anthropic’s cost guidance recommends pricing the workload’s costly tail. Keep their workflow totals visible instead of relying only on an average token count.

Model low, expected, and high cases for adoption, success rate, retries, review time, and maintenance effort. Those sensitivities show whether the proposed investment survives plausible operating conditions. They are assumptions in the ledger, not vendor performance guarantees.

Architecture / decision scenario

A financial-services operations team proposes Claude-assisted reconciliation of transaction exceptions. Every case requires analyst approval. The pilot handles 10,000 cases monthly and accepts 8,000 final reconciliations. All amounts are illustrative accounting assumptions, not Claude list prices.

Its monthly ledger includes $600 for inference across all attempts, $300 for retrieval, $400 for tool infrastructure, $500 for evaluation, $6,000 for analyst review and correction, $2,000 in allocated integration cost, and $1,200 for maintenance. Total cost is $11,000, or $1.375 per accepted reconciliation. Retry analysis attributes $100 of inference and $50 of tool costs to repeated attempts; those amounts are already inside the category totals.

The comparable baseline costs $1.60 per accepted case with the same approval scope. A lower-cost model saves $200 of inference but requires $800 more analyst correction. Its total rises to $11,600, or $1.45 per accepted case, so the architect retains the original candidate. If accepted outcomes fall to 6,000 with otherwise unchanged costs, the original path costs about $1.83 per accepted case and loses its cost advantage.

Finance reviews the launch funding separately from monthly amortization. The platform owner reconciles token estimates with the Admin API; operations validates review hours and accepted cases. The investment decision uses the complete workflow boundary rather than the attractive per-call price.

Common exam traps

• Presenting token spend as total cost of ownership.
• Dividing all costs by successful API calls instead of accepted business outcomes.
• Excluding failed attempts and then forecasting an unrealistically cheap workflow.
• Adding a retry subtotal twice when its inference and tool costs are already included.
• Treating workspace cost reporting as request-level business attribution.
• Assuming cache reads, writes, input, and output all have one price.
• Hiding launch funding, maintenance, or mandatory human approval behind amortized inference savings.

What to remember

Price every resource inside the workflow boundary and every attempt inside the case. Reconcile vendor charges, add people and lifecycle costs, and divide by accepted outcomes. Defend savings with consistent units, explicit allocations, and sensitivities that expose review burden and failure rates.

Knowledge check

References

• Anthropic, Optimizing for cost and intelligence
• Anthropic, Pricing
• Anthropic, Prompt caching
• Anthropic, Usage and Cost API