AI spend arrives through many doors: per-seat licences for coding tools, usage-based charges for premium models and agents, API tokens for product features, and compute for self-hosted models. Each behaves differently.
How the cost shapes differ
- Seats are predictable but easy to over-buy; unused seats are pure waste.
- Usage-based plans track value more closely but can spike — agents and long contexts consume far more tokens than autocomplete.
- API tokens for features scale with your users; small prompt changes can double the bill.
- Self-hosting trades token fees for GPUs, engineers and utilization risk.
Pricing models change
Vendors are still discovering their own economics, and they change plans. In June 2025 Cursor moved its Pro plan from request limits to compute-based limits that included $20 of frontier model usage per month at API pricing. On July 4, 2025 it published "Clarifying our pricing", saying the changes "were not communicated clearly, and we take full responsibility", and offered refunds for unexpected charges between June 16 and July 4. The lesson for buyers: model the cost of your usage pattern under each plan, and expect changes.
The FinOps basics for AI
- Visibility: tag all AI spend by team, use case and vendor; the gateway does most of this for API usage.
- Unit costs: cost per merged PR, per resolved ticket, per 1,000 feature requests.
- Budgets and limits: per-team budgets, per-key rate limits, alerts at thresholds, caps on agent run length.
- Optimization: right-size models (smaller models for simple tasks), cache repeated requests, trim prompts, batch where latency allows.
- Reviews: a monthly look at the top cost drivers and vendor plan changes.
Compare cost with value
A coding tool that costs a few hundred dollars per engineer per year is cheap if it saves even a few hours a month — but only if those hours turn into delivery. Tie cost to the scorecard's outcome metrics, not to time-saved surveys alone.