The Money Words: Why Your AI Bill Doesn’t Work Like Your Other Software

AI isn’t billed like the software you’re used to budgeting, and that’s been a running theme in a lot of the client budgeting work I’ve done. With traditional SaaS, the shift toward consumption-based pricing was actually something vendors chased for years — sell a company three or four seats and your revenue is capped by seat count, no matter how much they actually use the tool. AI providers skipped that struggle. Seat-based pricing showed up briefly, but the model settled fast into something purely consumption-based, and now the bulk of what you pay tracks usage, not headcount.

The basic unit of that consumption is the token — roughly a fragment of a word, sometimes a whole short word. Every model tokenizes what you send in and what it sends back, and you’re billed on both, usually at different rates, with input typically cheaper than output. One detail that surprises people: every new message in a conversation resends the entire prior conversation back into the model to generate the next response. That’s why input tokens tend to run much higher than output tokens over the course of a long session — you’re not just paying for what you type, you’re paying to resend everything that came before it, every single time.

Inference cost is what it actually costs a provider to run the model and produce your answer — distinct from training cost, which is the money and compute spent building the foundation model in the first place. Inference is the ongoing, metered cost of serving you a response every time you use the tool. Nobody in the industry is fully public about their real margins here. You’ll hear claims that a given lab is profitable on inference, then a caveat that the number excludes training costs and overhead. Take those claims with real skepticism — the actual unit economics of running these models profitably are still an open question industry-wide.

Token economics is the broader discipline of understanding what’s actually driving your bill up or down — which matters because pricing varies significantly by model tier and by provider, and it keeps shifting. One recent example: a provider cut its per-token price in a way that looked like a straightforward 40% discount, but paired it with a new tokenizer that uses more tokens to do the same work. Net effect: similar total cost, dressed up as a discount. If you’re not watching the mechanics, you’ll misread the number.

The practical fix is to stop budgeting AI like a flat SaaS seat and start budgeting it like a utility — a consumption forecast with a ceiling, not a fixed monthly number. In practice, I’d split AI spend into two buckets. One is general, ad hoc usage — people experimenting, trying things, using AI in ways you can’t cleanly tie to a specific task. Treat that like R&D spend you’re comfortable writing off, with a return that’s real but hard to quantify precisely. The other bucket is productionized use — a specific task, a specific workflow, something you can forecast token usage for and actually run an ROI calculation against. That second bucket is where real budgeting discipline pays off, and it’s worth re-baselining your cost tracking any time a vendor changes model version or pricing — the underlying unit economics can shift without your subscription line item changing at all.

Key takeaways

• A token is the basic billed unit — a fragment of a word, measured on both what you send in and what the model sends back.
• Every new message resends the full prior conversation, which is why input tokens usually outpace output tokens in longer sessions.
• Inference cost is the real, ongoing cost of running the model to serve your answers — separate from what it cost to train the model in the first place.
• Token economics means watching the mechanics behind a price change, not just the headline number — a “discount” paired with a new tokenizer can be a wash.
• Budget AI like a utility: split ad hoc R&D usage from productionized, forecastable workflows, and re-baseline whenever a vendor changes model version or pricing.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Money Words: Why Your AI Bill Doesn't Work Like Your Other Software