What's on the tab?

Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.

Summer 2015. I walk into a resto-bar on NYC's Upper East Side. Bottomless mimosas (and a forgettable brunch entree): $35. It was the best of times, it was the worst of times (for my liver).

Summer 2026. I walk into a club at the Encore Boston Harbor (cough, Everett, cough). Cover charge: $42.75. Drinks: expensive, and a la carte.*

Welcome to buying AI software in 2026. Bottomless brunch: tired. A la carte drinks: wired.

The Internet age of software was built on an assumption: once you built the product, each customer was virtually free. That one little detail shaped everything about pricing and packaging: seat-based licensing, subscriptions, unlimited usage, feature-based tiers. SaaS economics in a nutshell.

Bottomless brunch worked because most people didn't drink that much.

Then AI came in and threw all of those nutshells onto the floor like an over-served regular at a dive bar that gives out free peanuts.

Every prompt costs money. Every generated image has a cost of goods sold. Every agent workflow burns compute.

The response? Software didn't abandon subscriptions. It made you leave your card behind the bar and open a tab. Meet: subscriptions, plus usage-based billing.

This solved the vendor's accounting problem. It didn't solve the customer's product problem.

Everyone thinks this is a pricing problem. I think it's a translation problem.

*Believability aside, I may have fabricated that latter example.

I analyzed the pricing structures of 22 companies — from foundational models to AI-native startups to traditional SaaS incumbents.

Here's what I found:

The operators writing about AI pricing reached the same conclusion: AI broke one of modern software's oldest economic assumptions. (I wonder if the AI hype cycle-pilled SaaS execs saw this coming...)

Bain analyzed 30 traditional SaaS vendors introducing generative AI capabilities (excluding AI-native vendors) in October 2025 and found "about 65% have introduced a hybrid approach, layering an AI meter... on top of seat-based pricing."

Variable costs pushed vendors toward usage-based billing. Buyer demand for predictable budgets pulled them back toward subscriptions. Hybrid pricing emerged as the compromise.

The industry seems to agree on how to charge. It's still figuring out what to charge for.

Hybrid pricing still leaves one significant question: what belongs on the tab?

Every AI company is selling the same underlying commodity: compute.

They've all packaged it a little bit differently:

They're all buying GPU inference and they're selling something else. (And no, they aren't all just "AI wrappers.")

That's the product problem.

The infrastructure is similar. The customer mental model is where they diverge.

Every company shipping AI has to perform the same act of translation. Judging by the pricing pages, most are still on Unit 1.

Kyle Poyar, Bessemer, a16z, and just about every SaaS pricing expert agree that AI companies need to abstract away raw compute. The interesting question isn't whether to translate it. It's where to stop.

In looking across 22 companies, I realized they weren't inventing 22 different pricing models. They were speaking three different languages.

The language they chose tells you a surprising amount about the product they're trying to build.

Pay for how hard the AI worked.

The customer buys compute.

What the customer sees: tokens, premium requests, credits, reasoning.

Best fit: products where your customer thinks in compute. Developer APIs, technical buyers, variable workloads.

Case Study: Anthropic.

Anthropic charges for a base subscription with usage limits, then sells you "usage credits" when you exceed these soft caps. Anthropic assumes customers already understand compute.

Although the usage limits can feel opaque at times, this model makes sense because the power users — engineers, technically-minded PMs — think in tokens.

Tension: This ends up being the wrong language for many customers.

Notion bundles basic AI features (chat, generate, autofill, translate, meeting notes) in their subscriptions and sells "Notion credits" for premium features like Custom Agents and Workers. More credits are consumed when the agent reads more, performs more actions, uses stronger reasoning, or runs more often.

Notion is asking knowledge workers to think like infrastructure engineers. Nobody opens a note-taking app wondering how much inference their AI-generated status update is going to consume.

Verdict:

Notion's AI pricing feels like API pricing wearing a SaaS trench coat.

Pay for the work the AI performed.

The customer buys work.

What the customer sees: familiar units of work.

Best fit: products where customers naturally think in tasks rather than outcomes. Creative tools, project management, document management.

Case Study: Figma.

Figma's subscription tiers include a monthly allocation of AI credits, with the ability to purchase subscription credit packages or enroll in pay-as-you-go billing for overages.

Those credits translate into familiar tasks: make image, edit image, remove object, and more. Each is something a designer was already trying to accomplish long before AI showed up to the party.

This makes sense. They aren't forcing designers to think in GPU time. They're selling creative productive capacity.

Tension:

As AI capabilities expand, not all tasks are created equal. Work Proxies often introduce increasingly complicated exchange rates behind the scenes — one image generation might cost 10 credits while another AI task costs 40 — but those exchange rates always terminate at the task itself.

Verdict:

Figma asks designers to think about what they're trying to create, not what the model is doing behind the canvas.

The accounting underneath may be complicated, but the customer's mental model isn't. Credits disappear into the background, serving only as the exchange rate between compute and familiar tasks.

Pay for what the AI accomplished.

The customer buys results.

What the customer sees: business outcomes.

Best fit: products that can clearly map to measurable outcomes. Customer relationship management, sales lead generation.

Case Study: HubSpot.

HubSpot sells a SaaS subscription bundle for their customer platform software (sales, marketing, service, etc.) and HubSpot Credits.

HubSpot's AI spans customer support, prospecting, and data analysis — all bought with the same HubSpot Credits.

HubSpot gives sales and marketing teams a way to compare AI labor directly against human labor.

Tension: Even companies that aspire to Business Proxies often expose multiple pricing abstractions at once. Salesforce's Agentforce combines per-action credits, per-conversation pricing, seat licenses, and bundled enterprise editions — evidence that enterprise procurement often demands multiple ways to buy the same underlying AI capabilities.

Verdict:

HubSpot is pricing digital labor. Credits become an exchange rate for the cost of your future AI coworkers. (RIP, SDRs.)

HubSpot still uses credits as a shared currency, but unlike Work Proxies, the exchange rate doesn't stop at individual tasks. It continues all the way to business-facing results.

Intercom goes one step further, abstracting even the exchange rate away. Customers simply pay per successful resolution. The internal accounting almost certainly still exists — it just never leaves the billing system.

Just like Spain Men's National Team star, Lamine Yamal — AI credits are so hot right now.

Credit models have absolutely exploded among SaaS vendors, with a 126% YoY increase from 2024 Q4 to 2025 Q4, and an expected 114% further increase through mid-2027.

They aren't the innovation though. The innovation is choosing what credits represent. At their best, credits translate compute into the native tongue of the user.

Adobe's Generative AI Credits let them support different AI models, features, and compute costs behind a consistent customer experience. Customers don't think about the API request cost for the image gen model, they think about generating an image.

Every AI company starts with the same problem: inference has a marginal cost. Credits are one way of translating that cost into a customer-facing economic model.

Credits solve a vendor accounting problem. The product decision is choosing what the customer is buying.

Sometimes that's work. Sometimes it's results.

So you've chosen what customers are buying. What happens next?

In the words of Tupac "things change, and that's just the way it is... things'll never be the same."

It’s not just the product that changes, it’s who owns the discipline. Pricing is quickly becoming a product discipline.

The thing your customers buy becomes the thing your team is accountable for improving.

If your customers buy compute, your product competes on AI capability. Anthropic asks: how can we make every token more capable? Better models, better harnesses, better reliability.

If your customers buy work, your product competes on task completion. Figma asks: how can we help designers accomplish more work? Better workflows, better prototyping, better handoffs.

If customers buy results, your product competes on business outcomes. HubSpot asks: how can we drive better sales outcomes? Better customer insights, better lead qualification, better sales outreach.

Your roadmap naturally expands around the thing you're selling.

Traditional SaaS gives vendors time to recover from a bad experience. A frustrating interaction in February doesn't necessarily affect a renewal decision in November.

Usage-based AI compresses that feedback loop dramatically.

Retention happens every interaction. Every prompt. Every API call. Every generated image. Each is an opportunity to prove your value, or remind customers they have alternatives.

There isn't a single churn moment, usage simply evaporates.

Instead of asking "will they renew?" you start asking "why did usage stop growing?" That's an entirely different product problem.

AI businesses don't just lose customers, they stop getting invited to the party.

Subscription companies are laser focused on a handful of input metrics that dictate their success: customer acquisition cost, conversion rate, and retention.

AI product teams suddenly have a new game of buzzword bingo to play in their weekly business review*:

These aren't finance metrics. They're directly related to the product decisions you make.

AI pricing isn't just changing how companies are billed for software.

It's changing what product teams build, how they think about retention, and how they measure success.

Every AI company starts with the same underlying economics.

The winners will be the ones that translate economics from Wingdings into Helvetica.

Let's play: spot the bullshit buzzword. If you're the first one to guess correctly, I'll roast your company in the next issue.

Outcome-based pricing isn't a new idea.

I saw an early version of it over a decade ago at Deloitte. On one digital transformation engagement, the team experimented with Value-Based Billing. Rather than charging purely for time or a fixed implementation fee, part of our compensation was tied to reducing a retailer's marketing production cycle from 26 weeks to 20.

We ultimately got it down to 16 weeks.

It was memorable because it was unusual at the time.

Nearly every consulting engagement I worked on was billed by hours + cost ("time and materials") or fixed fee. The deliverables were agreed on upfront and no business outcome was tied to them other than perhaps "we launched the thing." Digital strategy. Data architecture redesign. Commerce implementations.

Value-Based Billing stood out because pricing outcomes is genuinely hard.

It required extraordinary trust (and copious team happy hours): executive sponsorship, constant communication, and a shared understanding of what actually drove the improvement.

AI skips the happy hours. When the software is the thing doing the work, the outcome measures itself.

For twenty years, software could afford to sell access. The economics made it silly not to.

AI has forced software companies to decide what kind of value they're actually selling.

The translation layer becomes one of the product's most important design decisions. What you choose to put on the menu shapes the workflows you build and, eventually, what the product becomes.

The most clever pricing model doesn't win. The company that wins is the one that best matches what it sells to how customers already think about the work.

Every AI company now keeps a tab. The question was never whether customers pay it — it's what shows up on the receipt.

Send me questions, feedback, and cocktail recipes:
[email protected].

_{Icons made by Icongeek26 from www.flaticon.com.}