Per-engagement token and cost reference · Anthropic API direct billing · September 2026
Token estimates: 1 token ≈ 4 characters of English text. A 3-hour session transcript at ~150 words/min ≈ 27,000 tokens. An 80-page PDF ≈ 40,000–60,000 tokens depending on density.
Prompt caching (Tools 04–07): Uploaded documents are sent first in the API message with a cache marker. After the first call in a session, document tokens are re-read at ~10% of input price. Cache TTL is 5 minutes — brief pauses between questions will not miss the cache; extended breaks may.
Diligence Coach note: As of this version, the API is only called once Target Context is added — clicking a question with no context shows it directly from the local question bank at no cost. The $3.83 full-run figure applies only once context is set and all 133 questions are generated fresh with documents attached. Cached localStorage responses from prior sessions are free — typical cost per engagement is materially lower.
Biz Dev Call Prep note: Web search is hard-capped at 3 queries per run (enforced by the API, not just prompted). If the first pass runs out of budget before finishing, one automatic retry occurs with a larger token budget and no further search — worst case ≈$0.43 instead of the typical $0.24.
SOW Generator note: The core builder — sections, sessions, pricing, phases, Word/PDF export — makes no API calls at all. Only the optional "Import from call transcript" and "Generate Email" features call the API, each gated behind its own cost estimate and confirmation prompt before running.
Observed vs estimated: A TDD Scorecard run with 3 large docs + 2 transcripts observed at $0.95 — consistent with the $1.17 estimate (variation due to actual vs assumed document sizes and output length).
Spend control: Set a monthly limit at console.anthropic.com → Settings → Billing → Spend limits.