Finance teams often treat AI spend as unpredictable. It isn't, structurally. Claude pricing is public and per-token: as of July 2026, Claude Opus 4.8 lists at $5 per million input tokens and $25 per million output tokens, Claude Sonnet 5 at $3/$15 (with introductory pricing of $2/$10 through August 31, 2026), and Claude Haiku 4.5 at $1/$5. Cloud marketplace list prices match these first-party rates. Your forecast is those prices times two things you can model: how many tasks run, and how many tokens each task consumes.
Two projection models that work
Volume-based projection is the simpler model and fits internal tools and pipelines. Measure the average tokens per task from real traffic, multiply by list price, then project task volume. Example with real list prices: a workflow averaging 1,500 input and 400 output tokens per task on Sonnet 5 at standard pricing costs about $0.0045 + $0.006 = $0.0105 per task. At 50,000 tasks a month that is roughly $525; a 10x growth plan puts the line item near $5,250 a month. The arithmetic is trivial — the work is keeping the tokens-per-task figure honest as prompts evolve.
Cohort-based projection fits user-facing features, where spend follows adoption rather than a pipeline schedule. Group users by signup month, measure tokens consumed per user per month for each cohort, and watch how that intensity changes as cohorts mature (early users often explore heavily, then settle). Total forecast = sum over cohorts of (users × tokens per user × price). This catches the pattern volume models miss: cost per user drifting upward as users adopt heavier features like long-document analysis.
Assumptions to document for finance
A forecast is only as auditable as its stated assumptions. Write these down explicitly:
| Assumption | Why it moves the number |
|---|---|
| Model mix | Opus 4.8 output tokens cost 5x Haiku 4.5 output tokens. A 10-point shift in routing changes the bill materially. |
| Sonnet 5 intro pricing expiry | Listed intro pricing ($2/$10) ends August 31, 2026; standard pricing is $3/$15 — a 50% step on both sides. Budget the post-August rate. |
| Tokenizer generation | Newer models (Opus 4.7+, Sonnet 5, Fable 5) use a tokenizer producing roughly 30% more tokens for the same text. Re-baseline per-task tokens when you upgrade models. |
| Cache hit rate | Cache reads bill at 0.1x the input price. A forecast assuming 80% cached input looks very different from one assuming none. |
| Batch share | Batch processing is billed at 50% of standard prices. Document what fraction of volume is batch-eligible. |
| Residency and endpoint premiums | US-only inference (inference_geo: "us") applies a 1.1x multiplier on eligible models; regional/multi-region endpoints on Bedrock and Google Cloud carry a 10% premium over global endpoints. |
| Output length drift | Output tokens cost 5x input tokens at every tier. Prompt changes that lengthen answers move cost more than input growth does. |
Calibrate against measured data, not estimates
On the first-party Claude API, the Usage and Cost Admin API gives you the ground truth to calibrate against: GET /v1/organizations/usage_report/messages breaks token consumption down by model, workspace, and service tier, and GET /v1/organizations/cost_report returns daily costs in USD. Re-fit your tokens-per-task and per-user intensity numbers against this data every planning cycle. Note that these endpoints are not available on Claude Platform on AWS, and on Bedrock, Vertex AI, and Foundry you calibrate from your cloud provider's billing and monitoring tooling instead.
Forecast the ceilings, not just the spend
Growth plans should also check platform ceilings. First-party organizations sit on usage tiers with monthly spend caps — Start ($500), Build ($1,000), Scale ($200,000), with a Custom tier arranged through the account team — and API usage pauses when the cap is hit. Rate limits (requests and tokens per minute) are also tier- and model-specific. If your forecast crosses a tier boundary mid-year, the planning conversation with your provider or account team should happen before the traffic does, not after a paused production system forces it.
Where to go next
For the pricing mechanics behind these numbers, see pricing explained and cost estimation. To connect spend to business value, continue with unit economics for LLM features, and set up anomaly alerting so the forecast gets a reality check every month.