Use actual invoices or dated list prices and label every unmeasured input. Together’s pricing page separates token-priced training, inference and GPU capacity: those units must not be mixed.
| Cost bucket | What to include | Common omission |
|---|---|---|
| One-off | Data cleaning, labelling, experiments, integration, migration | Buyer staff time and failed runs |
| Per workload | Calls, tokens, retries, routing, review | Double-processing before fallback |
| Capacity | GPUs, storage, networking, redundancy | Idle and peak reserve |
| Ongoing | Monitoring, incidents, supplier support, internal ownership | Updating tests after workflow changes |
| Exit | Export, rehosting, transition overlap | Egress and compatibility work |
Run low, expected and high volumes, plus a cheaper-API scenario and a higher-review scenario. Set the API-price change as a hypothetical stress input, not a prediction. If recurring savings are non-positive, simple payback is not attainable under those assumptions.
The supplied background’s $15,000–$40,000 feasibility and $50,000–$200,000 production ranges are unverified planning estimates. Do not substitute them for a quote or add overlapping ranges together. For an internal build, budget role-days multiplied by your own fully loaded daily rates, plus metered services, reviewer time and contingency approved by finance.