How to Think About Token Pricing: 4 Mental Models for AI Founders

Key takeaways

• A Hacker News discussion on token pricing surfaced recurring mental models rather than a single answer; the most practical framing treats LLM spend as COGS.
• Model 1, tokens as COGS: token costs belong in gross margin, giving every feature a unit cost.
• Model 2, tokens as a metered utility: you are billed per unit, so flat-priced products carry the usage risk.
• Model 3, subsidy-era pricing: today's list prices may sit below providers' true costs, so margins need buffers against repricing.
• Model 4, tokens as project capex: one-off agent bursts are investments judged on cost per outcome, not monthly bills.

Model 1: Why should tokens count as COGS?

Because they scale with usage, like hosting, not with headcount, like tooling. Once token spend sits in cost of goods sold, useful things happen: every feature acquires a unit cost, every user tier acquires a gross margin, and "can we afford a free tier" becomes arithmetic instead of vibes.

The practical consequence is caps. If each Pro user's included usage costs you, say, $6 at the median and $19 at p95, a $29 tier has a defined margin band instead of an open-ended exposure. Without the COGS framing, nobody ever computes that band.

Model 2: What does "metered utility" change about pricing?

You buy tokens on a meter. If you sell flat subscriptions, you have created a structural mismatch: variable costs under fixed revenue, and heavy users arbitrage the difference. The thread's sharpest framing: align your pricing axis with your billing axis, or hold reserves for the spread.

That does not mandate usage-based pricing for everyone. It mandates knowing your usage distribution and pricing the tail deliberately: caps, overage rates, or fair-use policies. Flat pricing with an unexamined tail is not simple pricing, it is unpriced risk.

Model 3: What if current token prices are subsidized?

A recurring thread theme: some list prices may sit below true serving cost while providers compete for share, so treat today's rates as a promotional era rather than a law of nature. The defensive move is a stress test: rerun your margins at 2x and 3x current token prices. If your business only works at subsidy-era rates, you have a countdown, not a moat.

The optimistic counter from the same debate: efficiency gains and competition keep pushing real costs down, so the subsidy may never need to end. Plan for the first scenario, enjoy the second.

Model 4: When is token spend really capex?

When a one-off burst buys a durable asset: a codebase migration, a bulk enrichment of your dataset, a one-time evaluation sweep. These should be budgeted per outcome ("this migration is worth $40K of tokens", as an illustrative decision) rather than absorbed into monthly COGS, where they distort every trend line. Founders who blend burst spend into run-rate spend misread both numbers.

Which mental model is right?

All four, applied to different parts of the product. Recurring per-user inference is COGS (model 1) bought on a meter (model 2) at possibly promotional prices (model 3), while occasional agent projects are capex (model 4). The failure mode the thread documents is not choosing the wrong model. It is pricing a product while holding none of them.

One-line takeaway: pick your mental model first, because your caps, tiers, and buffers all inherit from it. Then put numbers on it: Calcaas turns these models into actual per-tier margins, token costs, and provider comparisons you can act on.