Pricing AI products from first principles: seats, tokens or outcomes

Every request to a model burns compute that costs money, and that cost keeps falling. Here is what it means for seat, usage and outcome pricing.

Thesis: Software sold by the seat cost almost nothing per use, while artificial intelligence (AI) has a real, falling and unpredictable cost per use, so every AI price is a decision about who carries that cost: the buyer of a seat, the buyer of usage, or a vendor paid only for outcomes.

Start from zero

What must be true to sell an AI product at a profit? Five things.

Each request costs money. A model reads input tokens and writes output tokens, and both are billed. Anthropic lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and Claude Sonnet 5.5 at $2 and $10. Writing costs five times reading.
The cost of a given capability falls fast. Epoch AI found that the price of reaching a fixed benchmark score fell between 9 and 900 times a year, depending on the task; matching GPT-4 on doctoral-level science questions got 40 times cheaper each year. Anthropic’s Opus line went from $15 and $75 on Claude Opus 4 to $4 and $20 on Opus 5.5.
Tokens per task rise. Agents call a model many times per job and reason before answering. In April 2026, GitHub said agentic use had made Copilot’s request-based pricing “no longer sustainable”.
Value per task is uneven. A password reset and a saved cancellation can use similar tokens and be worth very different amounts.
Buyers need a bill they can forecast. Finance teams budget in seats and annual contracts.

Classic software as a service (SaaS) never faced the first fact: a seat cost about the same to serve, used once or all day. In 2020, Andreessen Horowitz (a16z) put the SaaS gross margin benchmark at 60% to 80% or more and observed that AI companies often ran at 50% to 60%. Bessemer Venture Partners’ State of AI 2025 report found its fastest-growing AI startups at about 25% gross margin, often negative, and its steadier “Shooting Stars” at about 60%, slightly below SaaS peers.

The primitives

Tokens, priced three ways. Input, output and cached input. On Claude Opus 5.5 a cache hit costs 5% of the input price, $0.20 per million, so a product that resends the same long instructions pays mostly cache prices.

Speed classes. The same Opus 5.5 model sells at three prices: $2 and $10 through Anthropic’s batch service, which returns results later, $4 and $20 as standard, and $8 and $40 in fast mode.

Credits. A prepaid unit that hides tokens. Salesforce sells Flex Credits at $500 per 100,000; a standard Agentforce action uses 20 credits, or $0.10. GitHub’s AI Credits are dollars metered on tokens at each model’s published rates.

Seats. A price per user per month. Salesforce’s Agentforce add-ons for Sales and Service cost $125 per user a month with unmetered employee usage; its $5 user licence has to be paired with metered credits.

Outcomes. Fin, an AI support agent that runs on Intercom’s helpdesk or others, charges $0.99 per outcome, at most one per conversation. A resolution means “no further help is requested after Fin’s last answer”; a conversation passed to a human without an outcome costs nothing. On other helpdesks there is a 50-outcome monthly minimum, and on Intercom’s own, human agents still pay $19 a seat. Sierra charges when its agent completes a task, such as resolving a conversation or saving a cancellation, and usually nothing otherwise; its post on the model names no rates. Salesforce lists Help Agent resolutions at $2.

Follow one request

A customer asks a support agent where her refund is. Assume the agent runs on Claude Sonnet 5.5 at list prices, makes four model calls and writes 2,000 tokens in all. These are illustrative numbers, not any vendor’s.

The app assembles 10,000 tokens of instructions, help articles and her order history.
The first call writes them to the cache at $2.50 per million: 2.5 cents. The model asks to look up the order.
The app queries the order system and appends the result.
Three more calls reread the cached 10,000 tokens, 30,000 in all, at $0.10 per million: 0.3 cents.
The 2,000 output tokens cost $10 per million: 2 cents. Inference so far: about 4.8 cents.
If she asks nothing more, an outcome-priced vendor bills $0.99. If she asks for a human, it bills nothing and still pays the 4.8 cents.
Priced as four Agentforce actions, the same conversation costs the buyer $0.40, solved or not. On a seat, it costs the buyer nothing extra.
At month end, if two in three conversations end in a billable outcome, inference per billed outcome is about 7.2 cents, 7% of the price. An agent that reasons and loops ten times as long spends about 72 cents, before people, tools or hosting.

Where the money is

Unit	Public example	Price
Seat	Agentforce add-on, Sales or Service	$125 per user a month, unmetered
Seat with credits	GitHub Copilot Business	$19 per user a month, including $19 of AI Credits
Usage	Agentforce Flex Credits	$0.10 per standard action
Outcome	Fin	$0.99 per outcome
Outcome	Agentforce Help Agent	$2 per resolution

Seats give the buyer a predictable bill and the vendor an unpredictable margin. A flat seat is overbooking: the vendor sells more allowance than the average user consumes, betting that most will not turn up. Agents make everyone turn up. Hence GitHub’s change on 1 June 2026: Pro still costs $10 a month and Business $19 per user, but each now includes the same value in AI Credits metered on tokens, and administrators choose whether to allow more usage or cap spending.

Usage keeps the vendor’s margin steady and passes falling token prices on, but the buyer pays for failed attempts and cannot forecast the bill without forecasting behaviour.

Outcomes move the risk to the vendor, and the upside with it. If the outcome price holds while the cost of capability falls 9 to 900 times a year, the saving becomes vendor margin. The danger sits in the two numbers the vendor controls least: tokens per task and resolution rate. Salesforce hedges by selling every unit at once, from $125 seats to $0.10 actions to $2 resolutions.

The airline parallel, honestly. Three parts transfer. Capacity: a graphics processing unit (GPU) idle for a second earns nothing, like an empty seat, and batch work at half price and fast mode at double are, in effect, off-peak and business fares for the same model. Segmentation: the fences are speed, model tier, allowances and who the user is. Willingness to pay: an outcome price anchors on what the buyer would otherwise pay a person, as a business fare anchors on the cost of a missed meeting. One part does not transfer. An extra passenger on a scheduled flight costs almost nothing, so a last seat can go cheap. An extra request always burns tokens, so an AI price has a floor. Unresolved conversations are the no-shows: the cost is spent and the revenue never arrives.

What is hard, and what I would build next

1. Bid prices for agents. Airlines accept a booking only if its fare beats the bid price, the expected value of keeping the seat. An outcome-priced agent needs the same rule for every step: before another model call or tool loop, compare its expected cost with its expected contribution, and hand off to a person when the next step costs more than it is likely to earn. The metric: inference cost per billed outcome, by type of request.

2. An outcome ledger the buyer can audit. “No further help is requested” is a definition the buyer has to trust. I would link every billed outcome to its transcript and credit it automatically if the customer returns with the same problem within a set window. A refund rule beats an argument about definitions, and the reopen rate becomes the number both sides watch.

3. Sell speed as fare classes. Anthropic already prices one model at half, standard and double. An application can pass that through: overnight reports at batch prices, live chat at standard, an “answer now” tier at a premium. Customers sort themselves by urgency, the same fence as advance purchase on an airline.