Pricing & Packaging Agent Products

9 min read

B5
Operation · Economics & ROI

Every agent price is an answer to one question: who carries the variance of the cost?

An agent's cost per user is not a number, it is a distribution with a long right tail, and the tail moves every time a model starts spending more tokens on a task. Seat pricing puts that variance on the vendor. Usage pricing puts it on the buyer. Outcome pricing puts the cost variance back on the vendor and makes the definition of an outcome the thing both sides are really negotiating. The 2025–2026 market has tried all three in public, and its reversals (Cursor's apology, Salesforce's move from per-conversation to per-action, the weekly caps on coding plans) all trace back to that one question. Choose a pricing model by deciding where the variance should sit and who can actually absorb it. Do not choose by copying what a competitor's pricing page looks like.

Step 1The SaaS margin intuition does not survive metered compute.

Classic SaaS priced a seat against near-zero marginal cost. Usage and margin were decoupled, so a customer who used the product all day cost about the same to serve as one who logged in once a month. An agent product reintroduces cost of goods. Every task spends model tokens, tool calls, retries and sometimes a human escalation, and the amount spent depends on how hard the task is and how autonomous the agent is allowed to be. Any pricing that ignores this is a SaaS model in an agent costume. It looks profitable until usage shifts, and usage always shifts, because the products that succeed are the ones people lean on harder.

The vendors say so themselves. When Cursor rebuilt its Pro plan in June 2025, the stated reason was that newer models use more tokens on longer tasks, so pricing by request count no longer tracked cost (Cursor, 4 July 2025). The same request, served by a better model, had become a more expensive request.

Step 2A worked example: the heavy tail breaks a flat seat.

Take an illustrative agent product sold at $40 per seat per month to 1,000 seats, so revenue is $40,000 a month. Fully loaded inference cost per seat is not uniform. It falls into three bands:

# month 1: illustrative cost distribution, $40/seat, 1,000 seats
light   700 seats × $3    =  $2,100
medium  250 seats × $15   =  $3,750
heavy    50 seats × $150  =  $7,500   # 5% of seats, 56% of cost
total                        $13,350  # inference margin 66.6%

# month 4: a new model doubles tokens on long agentic tasks
light   700 seats × $3    =  $2,100   # short tasks: unchanged
medium  250 seats × $30   =  $7,500
heavy    50 seats × $300  = $15,000   # each heavy seat now loses $260
total                        $24,600  # inference margin 38.5%

Three things in this example are worth taking away, and none depends on the particular numbers.

  • The median seat never moved. A dashboard that reports the typical user ($3) shows a healthy business in both months. The mean went from $13.35 to $24.60 because it is set by the tail, and the tail is where margin goes.
  • The vendor carried all of it. No customer's bill changed. That is what seat pricing is: the vendor has sold insurance against the buyer's cost variance, at a fixed premium, without pricing the policy.
  • Your best customers are your biggest losses. The 50 heavy seats are the most engaged, most likely to renew, most likely to evangelise, and each one costs $260 a month to keep. Under a flat seat, success and loss are the same cohort.

Anthropic's August 2025 move on Claude's subscription plans is this example in production. It added weekly usage limits on top of the existing five-hour limits for the $20 Pro and $100/$200 Max plans, effective 28 August 2025, citing users running Claude Code continuously in the background. It estimated the limits would affect fewer than 5% of subscribers (TechCrunch, 28 July 2025). Five percent of users was a big enough share of cost to justify a policy change across every plan.

Step 3Three models, three places for the variance to land.

  • Per seat means the vendor carries the variance. The buyer gets a predictable bill, and the vendor's defence is a quantity cap (rate limits, weekly limits, "fair use"), which is a price increase expressed in hours rather than dollars. It works when the cost distribution is narrow and the tail is small.
  • Usage or credits means the buyer carries the variance. Price tracks cost, so the vendor's margin is stable, but the buyer now pays for effort, including failed attempts and the vendor's own inefficiency, and gets a bill they cannot forecast. GitHub Copilot is the clean example: as of October 2026, Copilot Pro is $10 a month with 300 premium requests and Pro+ is $39 with 1,500, extra premium requests cost $0.04 each, and each model draws down the allowance at its own multiplier (GitHub Docs). The multiplier is the variance, passed through.
  • Per outcome means the vendor carries the cost variance again, but the price is attached to a result. The buyer pays only for what worked, and the vendor eats every failed and every expensive attempt. What the two sides are really agreeing on is the definition of an outcome. Sierra's December 2024 statement of the model: it is paid "only when we complete a task," unresolved and escalated conversations are "in most cases" not charged, and criteria for each outcome are agreed upfront (Sierra).

No model removes the variance. Each one moves it, and the question is whether the party you moved it to can see it, forecast it and survive it.

Step 4Moving variance onto the buyer is a contract change, so announce it like one.

Cursor's 2025 change is the cautionary case, and the lesson is not that usage pricing is wrong. The old Pro plan included 500 requests a month for external models, with Sonnet models counting as two. The new one included unlimited Tab and Auto-mode use plus $20 of frontier-model usage at API prices, with further usage bought at cost. By Cursor's own median-usage estimate, the $20 covered about 225 Sonnet 4 requests. The change took effect on 16 June 2025. On 4 July the CEO published an apology saying the change had been poorly communicated and had damaged trust, admitted it had not been made clear that "unlimited" applied only to Auto, and offered refunds for unexpected charges incurred over the previous three weeks (Cursor).

Read that through the variance lens. Before the change, Cursor carried the cost tail of a heavy Sonnet user. After it, that user carried their own tail, and the first they heard of it was a bill. The economics were defensible. The failure was that a transfer of risk reached the buyer as a surprise. The fixes Cursor shipped are the generic ones: a visible usage meter, a spend limit the buyer sets, a grace period, and advance notice of future changes.

If you are moving from seats to usage, the meter has to ship before the bill. A buyer who can see their consumption in real time is carrying variance they can manage. A buyer who learns about it from an invoice is carrying variance you hid from them, and they will treat it that way.

Step 5Under outcome pricing, the definition is the price.

Intercom's Fin is the most public per-outcome price. As of October 2026 it is $0.99 per outcome, charged at most once per conversation, and an outcome is counted when the customer confirms the issue is resolved, or the customer does not ask for more help after Fin responds, or Fin completes a workflow, including a handoff (Intercom pricing). The second and third clauses matter as much as the dollar figure. "Did not ask for more help" counts silence as success, and "including handoffs" counts a completed routing step as an outcome. Neither is unreasonable, but both are commercial decisions about where the line sits. Zendesk's August 2024 announcement charges for "issues that are resolved autonomously by AI," with a starter allotment included in every Suite and Support plan (Zendesk). Even there, deciding what counts as autonomous is the work.

The vendor's side of the same arithmetic, with illustrative costs:

# the cost of a billed outcome is the cost of every attempt that produced it
cost_per_outcome = loaded_cost_per_attempt / billable_rate

# $0.25 per attempt, 55% billable      → $0.45 per outcome (54% margin at $0.99)
# hard-case mix: $0.60, 40% billable   → $1.50 per outcome (a loss on every sale)
# same agent, "no follow-up" counts:   → $0.25 / 0.70 = $0.36

The last line is the point. Widening the definition of an outcome cut the cost of each billed outcome by a fifth, and nothing about the agent changed. That is why the definition belongs in the contract, with an audit path the buyer can check. A buyer who discovers the definition was generous will price that into the renewal.

Step 6Price the p90 month, not the average one.

Whichever model you choose, the variance you keep needs a buffer. The useful number is the cost of your 90th-percentile month, when retries, long-horizon runs and escalations spike together, not the blended average. A multiple over average unit cost is a buffer against that month, not margin padding, and the right size for it is something you derive from your own per-run cost distribution, not borrow from a benchmark. Autonomy widens the distribution. A tightly scoped agent has a narrow one. A long-horizon agent that can decide to spend 50× on a hard instance has a wide one, and whoever holds the variance under your pricing is effectively writing insurance on that tail.

Step 7Hybrids exist to split the variance, so split it on purpose.

The market has converged on composites, and each component is placed where one party can bear it:

  • Included allowance plus metered overage. Copilot's premium-request allowance and Cursor's $20 usage pool are both this shape. The vendor absorbs predictable volume, and the buyer pays the marginal cost of their own tail, ideally under a spend limit they set themselves.
  • Seat plus quantity cap. Anthropic's weekly limits are this shape. The buyer keeps a flat bill, and the vendor bounds the tail in hours rather than dollars. It is honest only if the cap is stated in terms the buyer can predict.
  • Finer usage units. Salesforce originally priced Agentforce at $2 per conversation, and in May 2025 added Flex Credits at $0.10 per action (20 credits; packs of 100,000 for $500). Its own example is an SDR interaction that cost $2 per conversation but 3–6 actions, $0.30–$0.60, under Flex Credits (Salesforce, 15 May 2025; Flex Credits explainer). A conversation is a coarse unit that hides a wide cost range. An action is a finer one that tracks cost, so the bill moves closer to the work actually done.
  • Platform fee plus outcome unit. The fee covers fixed cost and dampens the vendor's variance. The unit aligns the upside with delivered value. Sierra's own guidance adds consumption pricing for interactions like routing, where an "outcome" is not meaningful.

Step 8Choosing: match the model to who can see and survive the tail.

  • Seat if your cost distribution is narrow, or you can cap it with limits you are willing to state publicly. Watch the mean, not the median, and model what a model upgrade does to your heaviest decile.
  • Usage if your buyers are technical, can watch a meter, and would rather pay for what they use than subsidise others. Ship the meter, the spend limit and the notice before the price.
  • Outcome only where the outcome is objective, attributable to the agent, and verifiable by the buyer, and where you can survive the worst cost month at your actual success rate. If "resolved" is disputable, outcome pricing turns every hard case into a loss you owe and every generous definition into a dispute at renewal.

For the costing inputs, see fixed vs variable costs and forecasting agent spend. The pricing model is a risk allocation before it is a revenue model. Decide where the variance sits, tell the party holding it, and price the month it spikes.