.png)
.jpg)
Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
An idea is gaining ground in the pricing world right now. You will hear it at conferences, read it in newsletters, and see it hardening into frameworks and forthcoming books. The idea says that credit-based pricing is not a bridge to something better. Credits are the destination. They are the way to price the agent economy.
Here is what may surprise you: our problem is not with credits. Credits are a perfectly good metric for certain jobs, and we have recommended them to our own clients. Tokens too.
Our problem is with where the conversation starts. Any pricing discussion that opens with "which unit should we bill in" has already skipped every question that matters. Who is the buyer? What does the agent actually produce? Can anyone name what that output is worth? Starting with the metric is starting at the end.
Fixating on tokens and credits at the top of a pricing discussion is a losing move. We will say it more bluntly. It is a stupid person's game.
What a credit is actually anchored to
Strip away the packaging and look at what a credit is. A token is the vendor's raw unit of compute cost. A credit is a token with a markup and a friendlier name. Both are denominated in what the work costs the vendor, not in what the work is worth to the customer.
You can decorate credits with value math. Publish a table showing that a hundred credits gets you roughly ten processed documents. Build a calculator. Rename the units. The anchor does not move. The unit still lives on the cost side of the ledger.
We wrote about this dynamic in Monetizing Agentic AI. Inference-heavy products have real COGS exposure, and that exposure can push a vendor toward credits or tokens even when those meters do not match the value the customer receives. That is the honest description of why most credit systems exist. They are cost protection first and pricing second.
There is nothing shameful about cost protection. Every pricing decision has to survive the P&L. But four decades of value-based pricing doctrine rests on one instruction: denominate and communicate price in the customer's value terms. A credit does the opposite. It communicates the vendor's operations. When an entire movement builds its philosophy on a cost-side unit while borrowing the language of value, something has gone wrong.
The two honest jobs for credits
So when do credits genuinely earn their place? In our experience, two conditions.
Condition one: the output cannot yet be quantified into a value unit. Some agents produce work nobody can count. Exploration, drafts, dead ends, iteration. Or they produce countable work at a reliability level that cannot survive scrutiny.
Devin is the instructive case here. Cognition prices it through ACUs, a hybrid of platform fee plus compute-anchored usage units. That anchoring to cost is deliberate. Devin succeeds on a minority of complex tasks today, so pricing on output value would invite an ROI conversation its reliability cannot yet win. As reliability climbs, Cognition can layer output-quality tiers on top without changing the metric. That is a cost-anchored unit chosen through a real decision process, for a reason, with an exit path. We called it one of the best-designed pricing metrics in agentic AI, and we meant it.
Condition two: the product solves so many use cases that it starts behaving like a utility. A model API powers a thousand different applications. A horizontal agent does research on Monday, drafts contracts on Tuesday, and cleans data on Wednesday. No single value metric can describe that breadth.
At that point the product has stopped being software that delivers a nameable outcome and has become infrastructure. Nobody prices electricity per moment of illumination. You meter it, because the uses are infinite and the value is contextual. One infrastructure lead at a Series C AI startup put it perfectly: "We realized we were running a utility company but billing like a magazine subscription." When your product is genuinely a utility, a meter is the honest price.
Notice what both conditions have in common. Credits are what you pick when value pricing is not yet available, or when value is too diffuse to name. They are a considered fallback. A fallback is a fine thing to be. It is not a philosophy, and it is certainly not a destination.
Where credits fail
The failure mode is simple to describe. It is putting a utility meter on a product that is not a utility.
If your agent resolves support tickets, the value unit exists. It is a resolution. If it books meetings, the unit is a meeting. If it recovers chargebacks, the unit is a recovered dollar. Hiding a nameable value unit behind an abstract cost unit accomplishes exactly one thing: it transfers cost risk from the vendor to the customer and calls it pricing.
The market spent the last year running this experiment, and the results are public.
Cursor moved from request caps to compute-metered credit pools in June 2025. Users who had budgeted a day of coding watched the balance vanish in an hour. The backlash was severe enough that the CEO apologized in writing that July, saying "we didn't handle this pricing rollout well, and we're sorry," and issued refunds.
Replit shipped effort-based pricing the same summer. Bills arrived after the work, users paid even when the agent failed, and a July billing glitch hit roughly 6% of users. The company conceded the rollout fell short of its own standards. One reviewer wrote that the model made AI spend "feel less like software and more like a casino."
Lovable remains the most cited example of credit-burn unpredictability in the category. Across Reddit, G2, and Trustpilot, the recurring complaint is identical: one stubborn bug can eat a stack of credits that a manual fix would have resolved in minutes.
Then there is HubSpot, the most telling case because it moved in the opposite direction. HubSpot made credits its standard AI currency in mid-2025. A Customer Agent conversation burned about 100 credits, roughly a dollar, with top-ups near $10 per thousand. In April 2026, HubSpot repriced: $0.50 per resolved conversation for the Customer Agent, $1 per recommended lead for the Prospecting Agent. The value units were nameable all along. The moment HubSpot named them, the credits moved backstage.
Even Salesforce, the example credit advocates cite most, tells this story on close reading. Its May 2025 Flex Credits announcement leads with CIO cost anxiety, citing internal research that 90% of CIOs say managing AI costs limits their ability to drive value. The Flex Agreement lets customers convert user licenses into credits and back. That is procurement flexibility, and genuinely useful. But notice that Salesforce still quotes per-action and per-conversation prices in the market. The readable unit does the selling.
The direction of travel
The broader data points the same way. Intercom's Fin charges $0.99 per resolution. Sierra prices per successful resolution and, per CNBC in May 2026, raised at a $15.8 billion valuation with ARR reaching $150 million. Zendesk prices its AI on automated resolutions, at customer-reported rates around $1.50 to $2.00 each. Chargeflow takes 25% of each recovered chargeback and charges nothing on losses. EvenUp prices per completed legal demand package.
ICONIQ's late-2025 survey of roughly 300 software executives found outcome-based pricing rising from 18% to 23% of companies in six months. Kyle Poyar's 2025 survey of 240+ companies found seat pricing falling from 21% to 15% while hybrids jumped from 27% to 41%. Madhavan Ramanujam has argued that agentic products with clean attribution can capture 25% to 50% of the value they create, against 10% to 20% for classic SaaS.
None of these signals crown credits as the future. They show companies pricing the value unit wherever one can be named, and metering only where it cannot. That is not a trend toward any single metric. That is a market learning to make the decision properly.
You need a framework, not a favorite metric
Which brings us to the actual point.
In 2021, the pricing community declared usage-based pricing a game-changer. We never did. We called it a tool for the job, one whose value depended entirely on a company's strategy and context. The companies that treated it as a universal answer spent years unwinding the damage. Credits in 2026 are the same movie with a new currency. The industry keeps anointing metrics because anointing a metric feels like an answer. It is not an answer. It is a way to skip the work.
The work is a decision framework. Ours is the Agentic Monetization Spectrum, and it rates any agent on three dimensions.
Zero Human Ability. How much human involvement does the agent still need? When a human still does most of the work, per-seat pricing holds, because the human is the anchor. When the agent does the work, pricing has to track output or outcome. There is no human left to anchor to.
Operational Domain. How broad is the scope? A narrow agent is a tool. A medium one is a job function. A broad one is a department, and at the widest end it becomes the utility we described above. The metric has to match what the buyer pictures they are buying.
Output/Cost Curve. How does output value compare to compute cost? When they rise together, cost-based meters work fine. When output starts to outpace cost by ten or a hundred times, value-based pricing becomes possible. When the ratio goes exponential, pricing off cost is leaving the business on the table.
When we mapped five leading agents on these dimensions in Monetizing Agentic AI, they did not land on one metric. They landed on five different answers. Per-seat is right for Cursor. A cost-anchored hybrid is right for Devin, for now. Per-seat is defensible for Harvey because law firms budget by headcount, whatever the spreadsheet says. A flat fee is wrong for 11x. Per-outcome is right for Sierra, because it clears three specific bars: clear attribution, immediate measurement, and high value per outcome.
Read that list again. It includes a cost-anchored credit-style metric, chosen deliberately, that we praised. It also includes an outcome metric we called the best-aligned in the set, next to a warning that outcome pricing is not a goal every agent should chase. That is what a framework produces: different answers for different positions. A philosophy produces one answer for everyone, which is exactly why philosophies about metrics keep failing.
How to run the decision
Before anyone in your pricing meeting says the word credits, answer four questions in order.
- Can the customer name the unit of value? If a buyer can point at something the agent produced and say what it is worth, price that thing. Do not bury it.
- Who owns the outcome? High autonomy with clean attribution opens outcome metrics. A copilot assisting a human cannot honestly claim the outcome, so an output or activity meter is more truthful.
- Is the domain narrow or utility-broad? Narrow domains have nameable units. Genuine utilities deserve meters, and credits are a fine meter.
- What does the output/cost curve look like? Linear curves can live with cost-anchored pricing. Inflecting and exponential curves are telling you to move up the value chain, on a timeline your reliability can support.
The metric comes last. It falls out of the answers. Sometimes it will be credits, and when it is, use them with guardrails: no expiring balances, no hard caps without pay-as-you-go, and published math that ties units to something the customer recognizes.
The bottom line
We are not against credits. We are against coronations.
A metric is a tool for a job. Credits do two jobs well: protecting cost when output cannot yet be valued, and metering products that have genuinely become utilities. Outside those jobs, they transfer risk to the customer and dress it up as innovation, and the market has spent a year proving it.
So ignore anyone who opens the agent-pricing conversation with the unit of account, whichever unit they are selling. The metric is the decision, not the starting point. Do the work, and the metric will tell you what it wants to be.