Ask the same questions of every viable bidder:
- Which baseline model, date, settings and optimisation effort underpin your claim?
- Which cases failed, and who selected the evaluation set?
- Which work and costs are excluded from this price?
- What is the buyer expected to label, build or operate?
- Which artefacts can we export, under which licences and in which formats?
- Who responds to a critical incident and who approves a model update?
- What would cause you to stop at feasibility instead of recommending a production build?
For infrastructure-heavy proposals, require the matching service unit: training tokens, inference tokens and GPU-hours are different line items. Together’s published price structure illustrates that distinction; it is not a quote for any proposed engagement.