How much does AI actually cost? A practical guide to pricing.
Most cost surprises come from one of three patterns. Once you see them, AI cost forecasting becomes straightforward.
6 min read
AI is cheap until it is not. The teams that scale AI usage smoothly understand three pricing patterns. The teams that get surprise bills usually missed one.
Pattern one: per-seat pricing
Most enterprise AI products price per user per month. A $20-$60 per user range is typical for individual use, with $30-$200 for enterprise plans. ChatGPT Business, Microsoft Copilot, Google Gemini for Workspace, Claude Team — all per-seat.
This pricing is predictable. You know your headcount. You can model it.
This pricing also under-charges power users and over-charges occasional users. If half your seats use AI rarely, you are paying for unused capacity.
Pattern two: per-token API pricing
If you build directly on an AI API (Anthropic, OpenAI, Google AI), you pay per token used. A token is roughly a word. Pricing varies by model:
- Cheap models (Haiku, GPT-4o-mini, Gemini Flash): a few dollars per million input tokens, similar for output. - Mid models: $3-$5 per million input tokens, $10-$15 output. - Frontier models (Claude Opus, GPT-4o, Gemini Ultra): $15-$25 per million input, $60-$75 output.
For perspective: a 1,000-word essay is roughly 1,300 tokens. A long-context analysis of a 50-page document is maybe 50,000 input tokens. A weekly summary job running across a Slack workspace might be a few million tokens a month.
The variance per workload is enormous. A small chat tool can run for $30 a month. A document-classification pipeline processing thousands of documents can be $30,000.
Pattern three: cost spikes from poorly-bounded loops
The cost surprise most teams hit: a chatbot or workflow that loops or retries or has no token limit. Someone deploys it on Monday. By Friday, the bill is $10,000. By the next Monday, it is $50,000.
The fix is upstream:
- Cap max tokens per response. - Cap max retries per request. - Set per-day API budgets at the provider, and alarms at 50 percent and 80 percent. - Test in a small data set before turning it on at scale.
How to forecast cost for a real pilot
The honest method:
1. Sketch the workflow. 2. Estimate the per-call input and output tokens. 3. Estimate the number of calls per day. 4. Multiply through. 5. Add 30 percent for retries and exploration. 6. Watch the actual bill weekly for the first month.
Most pilots cost less than people fear and more than the most optimistic Twitter post. Build the forecast, test it for two weeks, adjust.
The hidden cost
The biggest cost is rarely the AI bill. It is the time spent on poorly-defined pilots that do not ship. That cost is invisible until you tally it.
The LearnTrainAI curriculum builds AI cost forecasting into the workshop directly.