Picture a gym that charges a flat monthly fee but quietly runs the numbers assuming most members never show up. The business model works beautifully until too many people start actually using the treadmills. This is not a metaphor for a fitness chain. It is a precise description of how flat-rate AI subscriptions are currently structured, and understanding it will change how you read every AI pricing page you encounter from here forward. ## The Assumption Buried in Every Flat Fee For roughly forty years, software pricing operated on a foundational truth: serving one more user cost almost nothing. As Utpal Dholakia wrote in The Pricing Conundrum, virtually every pricing strategy associated with SaaS offerings, from per-seat subscriptions and free tiers to volume discounts for large enterprise customers, was developed during a period when the incremental cost of adding a user was close to zero. That assumption shaped how founders modeled unit economics, how investors valued ARR, and how customers learned to think about what software should cost. A power user and a casual user were economically interchangeable from the vendor's perspective. Flat fees made sense because the cost curve was flat. AI inference is not flat. Every query, every generated image, every agentic task that runs autonomously on a user's behalf burns real compute that scales directly with how much the user actually engages. According to RevenueCat's analysis of subscription app economics, AI features introduce variable costs that scale with user engagement, directly disrupting the near-zero marginal cost model that traditional subscription apps were built on. The more a subscriber uses the product, the more it costs to serve them. That is structurally the opposite of how SaaS economics were designed to work. ## Why Vendors Set Prices That Can Turn Against Them So why would any rational company price a product this way? The answer, as Marc Nager noted in a widely-circulated analysis of AI economics on LinkedIn, is that the cost burden is being actively subsidized across the major AI players. Brad Feld's framing, cited in that same discussion, captures the structural tension cleanly: AI might be faster, but it is not cheaper, and there is a building debt underneath the pricing that consumers are not seeing. Flat-rate pricing is not naivety. It is a calculated bet that most subscribers will use the product lightly, generating revenue that cross-subsidizes the minority of heavy users. It works until usage patterns shift. The shift is already visible. Agentic AI use cases, where a model executes multi-step tasks autonomously rather than answering a single query, drive far heavier compute consumption per session than conversational use. As Lago's analysis of AI pricing models makes clear, monitoring usage and choosing the right pricing model is genuinely high-stakes for any AI business: one wrong assumption about engagement depth can collapse the unit economics that the whole product margin depends on. The editorial team at Aria Systems put it more directly, arguing that flat-rate subscriptions are structurally incompatible with AI product economics over time. ## What the Pricing Redesign Actually Looks Like The industry is not standing still. Dholakia's Pricing Conundrum piece, published in May 2026, documents a full generation of new structures being developed in response: tighter free tiers, usage caps, metered billing, outcome-based fees, labor replacement benchmarks, and frameworks for valuing AI agent labor. Bessemer Venture Partners has published a detailed AI pricing and monetization playbook mapping how vendors can align price structures with the actual cost of delivering AI value. The common thread across all of these approaches is a move away from the flat-fee assumption toward structures that let revenue scale with the compute being consumed. TechCrunch's June 2026 reporting on Google's moves in the AI subscription market signals that even the largest players are actively repricing and repositioning, which means the competitive landscape is in genuine flux. For anyone building a product on top of AI infrastructure, RevenueCat's recommendation is operationally concrete: model AI usage against your ARPU and retention metrics, route requests to cost-efficient models where possible, reuse generated outputs rather than regenerating them, and gate advanced AI access behind paid tiers that reflect the actual cost of serving heavy users. These are not theoretical best practices. They are the mechanics of surviving the transition from flat-fee assumptions to a world where utilization actually matters. ## What to Watch and What to Build Toward The pricing redesign Dholakia describes is still early. Most of the industry conversation, as he notes, has focused on the seller's perspective, with almost no serious work done on how customers will actually perceive and evaluate the new structures. That gap matters enormously for anyone building AI-powered products: you can design a metered or outcome-based pricing model that is economically rational and still lose customers who were trained by forty years of flat-fee SaaS to expect predictable monthly bills. The customer-perception half of this problem is the one that will determine which new pricing structures actually stick. The constructive read here is that this moment is genuinely instructive. Every AI subscription you evaluate, whether as a user, a builder, or an investor, now comes with a hidden question worth asking: what utilization rate does this price assume, and what happens to the business if actual usage looks different? The teams that internalize that question early, and build pricing architectures that can flex with real-world engagement patterns, are the ones most likely to be standing when the subsidized-growth era of AI pricing ends. ## Sources - Subscription App Economics: The Hidden Cost of AI Features

Sources