AI Unit Economics: Pricing & Margins for AI Services

AI unit economics break the classic SaaS playbook in exactly one place — cost of goods sold. Traditional software is built once and served to the next customer for almost nothing, which is why mature SaaS runs 80-90% gross margins. An AI service cannot copy that trick: every query spends real inference, so gross margins land closer to 50-60%. This guide is a working framework for the COGS levers, the cross-vendor caching math, and a pricing-model stress test you can run on your own numbers.

The stakes are concrete and immediate. A flat monthly price that looks healthy on a blended average can be quietly negative on your heaviest 10% of users, because an agentic loop consumes far more tokens than a single chat turn. GitHub watched exactly that happen to Copilot and moved every plan to usage-based billing in 2026. Pricing an AI product or an AI-powered service without modelling the inference underneath it is not optimism — it is exposure.

What follows is the framework we use when we re-price AI offerings: why the margin ceiling is lower than SaaS, the three levers that actually move COGS, a side-by-side of how the three frontier vendors price caching, a worked stress test showing where each pricing structure breaks, and a read on what the market is converging toward. Where a number is illustrative, it is labelled illustrative.

The defining feature of software-as-a-service is that the marginal cost of one more customer is close to zero. You build the product once and serve it to the thousandth user for roughly what it cost to serve the tenth. That is the entire reason mature SaaS runs 80-90% gross margins, and it is the assumption every classic SaaS pricing model quietly rests on.

AI breaks that assumption at the point of cost of goods sold. There is no build-once-serve-infinitely trick for inference: every query sends tokens through a model, and every token has a price. The cost does not amortise away with scale — it recurs on the next request, and the next. A product can look profitable in a blended average and still be losing money on its most engaged users, because those users are the ones generating the most tokens.

Three things keep AI COGS stubbornly non-zero, and each one is a place where a naive flat price springs a leak.

Put the benchmarks next to each other and they agree. Bessemer Venture Partners’ AI pricing and monetization playbook, fetched from the live page, states AI companies see 50-60% gross margins against 80-90% for traditional SaaS. ICONIQ Capital’s State of AI snapshot, surveying roughly 300 software executives, puts the average AI product gross margin at 52% for 2026 — up from 41% in 2024, so the trend is improving, but still a full tier below software.

The composition matters as much as the headline. Within ICONIQ’s sample, companies that combine model-layer and product-layer innovation — what the report calls balanced differentiation — report the highest margins at 53%, while pure application-layer companies reselling someone else’s model sit at 45%. The closer you are to just passing through a frontier API, the thinner the margin you keep.

Here is the counterintuitive finding that should reshape how you plan. The instinct from software is that COGS shrinks as a share of revenue with scale. ICONIQ found the opposite for inference: model-inference cost climbed from 20% to 23% of total spend as products moved from early to mature stage, while talent dropped from 32% to 26%. As an AI product succeeds, the meter runs faster, not slower — usage grows, inference grows with it, and the line you most wanted to dilute becomes a bigger part of the bill. Planning for margins to expand automatically with scale is planning on the wrong curve.

If the margin ceiling is structural, the response is engineering: drive the cost-to-serve down so a defensible margin survives at your price. Three levers do most of the work, and the largest one is free — it is an architecture decision, not a discount you negotiate.

Routing is where most teams leave the most money on the table. The reflex is to run everything on the best model, but the price spread between a cheap and a frontier model from the same vendor is routinely 5x or more, and a large share of real traffic is routine enough for the small model to handle. For the engineering detail on the caching lever, our prompt-caching engineering guide walks through breakpoints and cache-life mechanics, and our guide to right-sizing model spend covers the gate-by-gate playbook for cutting an AI bill. A worked example of the routing lever in practice — near-frontier coding performance at a fraction of the price — sits in our Claude Sonnet agentic-coding cost breakdown.

Caching deserves its own table because it is the rare cost lever where the whole frontier has independently converged. We fetched each vendor’s live pricing on June 30, 2026 and laid the mechanics side by side. The read discount is the same story everywhere — roughly 90% off for re-used input — while the way each vendor charges to create and hold a cache differs.

The convergence is the interesting part. When three competing labs independently price the same lever the same way, it stops being a promotional discount and starts looking like a structural property of how transformers serve repeated context. The practical takeaway: if your agent re-sends a stable prompt — and almost every agent does — caching is not an optimisation you get to later, it is part of the cost model from day one. Just price the write premium and any storage fee into the math: Anthropic’s 1.25x write and Google’s hourly storage charge mean caching pays off on reuse, not on a single call.

Most coverage repeats the 50-60% headline without showing why a flat price can go underwater on its heaviest users while looking profitable on average. So here is the math, worked. The numbers below are illustrative, not market data — but they are built on real per-token prices so the mechanics are exact.

Assume a cost-to-serve of $0.04 per agent task: a 3,000-input plus 1,000-output-token task at Claude Opus 4.8 list prices ($5 and $25 per million tokens) is (3,000 × $5 ÷ 1M) + (1,000 × $25 ÷ 1M) = $0.015 + $0.025 = $0.04. Three monthly intensities: a light user runs 100 tasks ($4 cost), a median user 500 tasks ($20), and a power user on long agent loops 2,500 tasks ($100). Now watch what three pricing structures do to gross margin across those users.

Read the power-user column first. The flat per-seat plan posts a healthy 86.7% margin on a light user and a comfortable 33.3% on a median one, then collapses to roughly negative 233% on the power user — you are paying $100 to serve someone who pays you $30. Pure usage-based pricing holds a flat 50% margin across every user by construction, but it gives up the predictable revenue floor that customers and finance teams both like. Hybrid threads the needle: 73.3% on the light user, 42.9% on the median, and 48.7% on the power user — never negative, never wild, with a base that guarantees a floor and overage that follows cost.

The forward read is that the industry is walking down exactly this table. As agentic usage spreads, the power-user column stops being a tail case and becomes a meaningful slice of the base, which is precisely why flat-fee AI products keep converting to usage and hybrid models rather than the other way around. The structure that survives contact with heavy agentic users is the one with a cost-following component built in. To build the column on the left — an honest cost-to-serve per task and per user — see our per-task, per-user cost framework; you cannot price what you have not measured.

The theory shows up in live pricing pages. The clearest case study is GitHub Copilot, which moved all plans to usage-based billing effective June 1, 2026, replacing flat request allotments with metered GitHub AI Credits charged on actual token consumption. Base subscription prices were left unchanged; what changed was that consumption now meters. GitHub’s own stated reason is the exact mismatch this whole article is built on.

That is a flat-fee vendor saying out loud that flat fees stop working once one user’s session can cost orders of magnitude more than another’s. The rest of the market is improvising different answers to the same problem.

For an agency or a software team putting a price on AI-powered work, the lesson is not to copy whichever vendor is loudest. It is to start from your own cost-to-serve, then choose the structure whose risk profile matches how variable your usage is. Bessemer’s framing is the right one: your charge metric is a strategic statement, not just a billing decision — tokens suit technical buyers but confuse everyone else, and outcomes align value but force you to absorb cost variability. Pick deliberately.

This is exactly the work we do when we re-price an AI offering: model the real cost-to-serve, apply the routing and caching levers until a defensible margin survives, then choose the pricing structure that matches the client’s usage profile rather than a template. It is where our AI transformation engagements begin, and it is the difference between an AI-powered service like CRM automation that compounds margin and one that quietly loses money on its best customers. Price the COGS first; everything downstream follows from it.