top of page
Digital-Price-Works-logo

Priced Before It Was Understood: What AI's Early Pricing Failures Are Teaching Us ___________________________

  • Writer: Jorge Rodriguez
    Jorge Rodriguez
  • 4 days ago
  • 6 min read

Imagine a restaurant that offers an all-you-can-eat buffet. For most diners it works — they eat a reasonable amount, the kitchen covers its costs, everyone leaves satisfied. Then a table of heavy eaters arrives and doesn't stop. The catch: the kitchen is cooking every bite to order. The cost of serving that one table begins to exceed what the entire section is paying. The restaurant is full, revenue looks fine, and the economics are quietly breaking.

This is, in compressed form, what happened to several prominent AI companies between 2023 and 2025. The product was unlimited AI capability at a flat price. The kitchen was cooking every token to order. And a small number of heavy users found they could eat far more than the price had ever accounted for.


The cases

Cursor sold a flat $20-a-month Pro plan with approximately 500 fast requests per month. On June 16, 2025, the company switched to a $20 credit pool billed at underlying API rates. The effect was immediate: newer models cost significantly more per token, and some users burned through their entire monthly allocation after two or three complex coding prompts. Twelve days later, the plan description was quietly changed from "Unlimited" to "Extended." On July 4, CEO Michael Truell issued a public apology — "We recognize that we didn't handle this pricing rollout well and we're sorry" — and offered refunds for the period June 16 to July 4. The economics behind the change were straightforward: multiple industry reports confirmed that margins on AI coding tools were "either neutral or negative," and Cursor was reportedly spending close to 100% of revenue on AI costs despite reaching $1 billion in annual recurring revenue by November 2025. (Sources: TechCrunch, July 7, 2025; FinTech Weekly, July 12, 2025; TechCrunch, August 7, 2025)


Replit launched its AI coding agent in September 2024, and revenue surged from $2 million annualized in August 2024 to $32 million by February 2025. Gross margins hit 36% in February — then fell to negative 14% by April, when the company launched a more autonomous version of its agent. The new agent ran longer, made more LLM calls, and consumed far more compute per session. The pricing model it inherited was built for the old product's cost profile; applied to the new one, it produced losses at scale. As one detailed industry analysis noted: "Replit's margin collapse was not a failure of execution. It was a failure of pricing architecture applied to a product with costs that SaaS models were never designed to handle." (Sources: Aakash Gupta, Medium, March 4, 2026; The Information, August 2025; TechCrunch, August 7, 2025)


Salesforce Agentforce launched at $2 per conversation — simple, easy to budget, easy to sell. Two problems emerged quickly: customers were confused about what counted as a "conversation" in complex, multi-step scenarios, and the rate was charged regardless of whether the agent resolved anything. Adoption was tepid; Salesforce's own figures showed only 3,000 paid deals in the first two quarters. On May 15, 2025, Salesforce introduced Flex Credits at $0.10 per action. By late 2025, a third model — per-user licensing at $125+ per month — was added. As of early 2026, all three pricing models run simultaneously on the same product. (Sources: Salesforce press release, May 15, 2025; Aquiva, Q3 2025; SaaStr, February 16, 2026)


Anthropic disclosed, in the context of industry reporting on AI gross margins, that on its original Claude Code pricing structure, the company was losing tens of thousands of dollars per month on a single user subscribed to a $200 plan. A small number of power users were consuming compute at a rate that erased the margin from hundreds of regular subscribers. (Source: Tanay Jaipuria, "The State of AI Gross Margins in 2025," September 2, 2025 — tanayj.com)



This isn't the first time

AT&T launched the iPhone in 2007 with a $30-a-month unlimited data plan. It worked because AT&T's own data showed 65% of users consumed 200MB or less per month, and the average iPhone user consumed just 251MB (Bernstein Research, late 2009). The flat rate was priced against average behavior — and average behavior was light. Then streaming arrived, then video calls, then gaming. In June 2010, AT&T became the first major US carrier to eliminate unlimited data for new customers, drawing backlash that, in tone and intensity, sounds familiar.


By 2017, competitive pressure forced all four carriers back to unlimited. Verizon's first full quarter with its new unlimited plan saw wireless margins contract from 47.8% to 45.8%. By Q2 2025, the EBITDA spread between the highest- and lowest-margin carrier had compressed 89% from 2020 levels — the sustained cost of competing on unlimited in a capital-intensive business. (Sources: Fortune, June 4, 2010; CBS News, June 2, 2010; Fortune, April 4, 2017; LightReading, February 20, 2026; Verizon Q3 2025 earnings)

The carriers stabilized through tiered "unlimited" plans that throttle above a usage threshold, good-better-best architecture that segments heavy users into higher-priced tiers, and sustained infrastructure investment that brought marginal cost per gigabit down over time. None of it was quick. All of it was resisted initially because it complicated a simple offer.


What pricing professionals should take from this

SaaS pricing assumes near-zero marginal cost. AI does not. Every user is a cost event. The heaviest users are the most expensive to serve. That single structural difference breaks flat-rate pricing every time, and will continue to until pricing models are built for AI's actual cost structure rather than borrowed from software.

The billing unit matters more than the price. Cursor priced on requests. Salesforce priced on conversations. Neither mapped to what the customer actually cared about. When the unit is wrong, changing the number doesn't fix it — the unit itself has to change. Pricing models that have stabilized in AI share one property: the unit is something both parties can verify.

Instrument before you price. Every case above involved pricing announced before the company could accurately meter its own production costs. Metering infrastructure is a pricing prerequisite, not a post-launch addition.

A new capability is a new pricing problem. The Replit case makes this exact point: a more autonomous agent is not the same product at a higher performance level. It has a different cost structure and requires pricing derived from that structure, not inherited from its predecessor.



The tab has arrived

The mobile industry faced the same table of heavy eaters, and its answer was not to throw them out — it was to redesign the menu. Tiered plans let heavy users keep eating at a higher price; throttling slowed service rather than cut it off; infrastructure investment brought the cost of each bite down over time. None of it was painless, and all of it required admitting that "unlimited" had never quite meant what it said.

AI pricing is at the same inflection point. The kitchen is real, the costs are real, and the heavy eaters are not going anywhere. The companies that learn to price what they actually cook — not just how many diners sat down — will be the ones still operating when the next wave arrives.

Sources



 
 
 

Comments

bottom of page