Freeze the workload before multiplying prices

Write down the feature and the unit source:

Field Buyer record
Business outcome and feature boundary
Provider and exact model/version
Requests per period
Input tokens per request
Cached input tokens per request
Output tokens per request
Tool, image, audio, search, storage, or embedding units
Low, expected, and peak assumptions
Unit source and access date
Workload owner

Do not fill these cells with an optimistic average because the request count is easy to guess. If the supplier cannot show how the units are measured, record that as a quote risk and ask for a metering plan.

Use a dated rate snapshot

The evidence packet records one illustrative standard paid-tier row per provider. Rates are volatile, model-specific, and qualified by each source:

Provider/model Base/uncached input per 1M Cached input per 1M Cache write per 1M (5m / 1h) Output per 1M Important qualification
OpenAI GPT-5.6 Terra $2.00 $0.20 $12.00 The official model comparison page has separate tool, batch, and capability details.
Anthropic Claude Sonnet 4.5 $3.00 $0.30 cache read $3.75 / $6.00 $15.00 Cache-write values are the 5-minute and 1-hour TTLs; endpoint choices need separate treatment.
Google Gemini 3.5 Flash $1.50 $0.15 context cache $9.00 Context-cache storage, grounding, and non-token charges are separate.

See the Anthropic pricing page and prompt-caching documentation, and Gemini API pricing for the provider-specific definitions. This table is a dated planning input, not a promise that these prices will remain unchanged or apply to your account.

Apply the formula transparently

For a simple token scenario, record disjoint usage categories once. In the worksheet, “uncached input” already excludes cache reads and cache writes; do not subtract cached tokens again:

token estimate = (
  uncached input × base input rate
  + cached input × cached-input rate
  + 5-minute cache-write tokens × 5-minute cache-write rate
  + 1-hour cache-write tokens × 1-hour cache-write rate
  + output tokens × output rate
) ÷ 1,000,000

The worksheet's synthetic test uses 600,000 uncached input tokens, 200,000 cached input tokens, and 100,000 output tokens against rates of $2.00, $0.20, and $12.00. That arithmetic produces $2.44. A separate test prices 5-minute cache-write units; neither fixture is a forecast for a customer. Run it with:

npm test --prefix sites/dealdigg.com/evidence/P94

Then add other charges in separate columns. Do not hide tool calls, images, storage, hosting, support, taxes, currency conversion, or supplier margin in a single “AI cost” number.

Model low, expected, and peak scenarios

Keep three rows rather than one false precision point:

Scenario Usage source API estimate Other metered charges Hosting/support Total Decision owner
Low Observed or explicitly assumed
Expected
High/peak

The buyer should define what happens when the high case arrives: alert, rate limit, queue, degrade to a cheaper model, pause a feature, or approve more spend. The cap owner and the operator who can take the action belong in the quote review, not only in an engineering backlog.

Reconcile usage with the invoice

After launch, compare the provider's usage or cost record with the invoice for the same period. Keep the model/version, unit counts, rate snapshot, credits or discounts, other charges, and unexplained difference. A mismatch is a change or investigation item, not a reason to silently adjust the worksheet.

Record rate changes and workload changes in a dated log. If a provider changes the model, tokenizer, cache-read or cache-write TTL terms, or tool price, re-run the low/expected/peak scenarios and ask the commercial reviewer whether the quote or support scope changes. The supplier's margin and any fixed maintenance fee should remain separate from provider consumption.

Commercial stop condition

Do not accept an AI cost assumption as quote-ready when the workload owner, measured unit source, exact model, cap owner, alert response, invoice reconciliation path, or rate-review date is blank. This worksheet helps expose the commercial questions; its cap result is a planning control, not provider billing enforcement. It does not set a contract price, guarantee a future bill, or provide legal advice. Have the buyer, FinOps, commercial owner, and qualified counsel review the actual agreement where appropriate.