Skip to main content

​
Evidence hierarchy

Use evidence in this order:
  1. Provider-billed (actual) costs for financial totals.
  2. ThinkPlatform allocated provider actuals for project, user, Agent, task, and model attribution.
  3. Runtime measured estimatedCost for diagnostic detail.
  4. Token/request/duration proxies when direct cost is unavailable.
Provider actuals and runtime estimates can differ because of reporting delays, currency conversion, discounts, service tiers, rounding, cached-token treatment, tool charges, credits, or incomplete attribution. Reconcile the differences rather than combining the figures.

​
General formula

For each provider, model, service tier, and context tier:
Then:
Do not assume all providers expose every component or that cached tokens are separate from total input tokens. Confirm field semantics before subtracting or adding token categories.

​
Metrics to calculate

  • Cost share: slice actual cost / project actual cost.
  • Activity share: slice requests / project requests.
  • Effective cost per request: slice actual cost / slice requests.
  • Effective cost per million input/output tokens: use only when token coverage is adequate and label it as blended.
  • Cache ratio: cached input tokens / input tokens only when input tokens include cached tokens.
  • Output amplification: output tokens / input tokens.
  • Reasoning share: reasoning tokens / output tokens only when the provider documents reasoning as part of output.
  • Daily execution rate: requests or tasks / active days.
  • Schedule overhead: failed, retried, duplicated, or overlapping executions × representative task cost.
  • Concentration: cost share of top model, Agent, task, user, and day.
Always handle zero denominators and disclose low token coverage.

​
Current pricing verification

Pricing is time-sensitive. At analysis time, verify the specific provider, model snapshot, region, service tier, and context length using official sources: Record the URL and access date in the report. Do not copy a price from this skill.

​
Pricing distinctions that change conclusions

  • Input, cached input, cache writes, and output can have different rates.
  • Long-context tiers can change rates above provider-specific thresholds.
  • Standard, Batch, Flex, priority/fast, and provisioned capacity have different economics.
  • Some built-in tools and storage are billed separately from model tokens.
  • Azure Standard token billing and provisioned throughput unit/hour billing require different utilization analyses.
  • A cheaper model that increases retries, output, or tool loops can cost more per successful outcome.

​
Prompt caching interpretation

Caching helps only when requests are eligible and share an exact stable prefix under the provider’s rules. Place stable instructions and common context before variable user/task content when platform behavior permits. Do not treat low cache usage as waste when:
  • requests are below the provider’s cache eligibility threshold;
  • prompts legitimately differ at the beginning;
  • traffic is too sparse or separated beyond cache lifetime;
  • the provider/model/tier does not support the expected cache behavior.

​
Currency and billing period

Keep provider currencies separate unless ThinkPlatform supplies a documented conversion. State the exchange-rate source and date when converting. Align analysis dates with provider reporting timezone and note that very recent days may be incomplete.