Introduction: The Seismic Shift

The $300 billion SaaS market in 2025 is not experiencing disruption—it’s experiencing bifurcation. On one side, traditional productivity applications like Notion, Trello, and Evernote are quietly retreating their free offerings and tightening collaboration limits. On the other, AI-native applications are pricing their services in ways that would have baffled business software executives just five years ago: by the action, by the reasoning step, by the token consumed.

This seems less like an adjustment and more like a fundamental reckoning between two irreconcilable business models. But AI applications can’t run on this model. Or rather, they can, but it destroys their unit economics. The variable cost of delivering AI—whether it’s the computational infrastructure, the API calls, or the specialized silicon—scales directly with usage. Charge per seat, and you’re betting that users sit idle most of the time. That bet is increasingly losing. What’s emerging instead is a market split into two distinct worlds, each with its own pricing logic, retention profile, and user psychology.

Understanding this split has become essential for anyone navigating the 2026 software landscape, whether you’re an investor evaluating AI startups, a product leader defending your pricing strategy, or a CTO trying to justify another piece of subscriptionware in your stack.

From Seats to Actions: The Fundamental Divergence

The economics of the per-seat model rely on a simple assumption: users are interchangeable units. If a user costs nothing to serve once the software is built, the marginal cost of adding another seat is nearly zero. Charge $15 a month per seat, and your gross margin approaches 95% as you scale. It’s a beautiful model for businesses that assume limited variable costs. For decades, it worked.

But AI breaks this. Every interaction with an AI model consumes computational resources—GPU cycles, memory bandwidth, specialized silicon. These have hard costs that scale directly with usage. Unlike a collaboration tool where a tenth user barely increases server load, an AI application’s infrastructure cost per user can vary by 100x depending on usage intensity.

This is why AI application pricing has inverted the traditional SaaS model. Instead of paying per employee, companies now pay per outcome or per unit of computation. Intercom’s Fin AI charges $0.99 per resolved customer support case. Salesforce introduced Flex Credits in 2025—$0.10 per AI "action." This isn’t licensing anymore; it’s paying for results.

This creates an entirely different pricing psychology. Traditional SaaS asks: "How many people do you want to help?" AI SaaS asks: "How much value do you want to extract, and how much are you willing to spend?" Meanwhile, legacy productivity apps have doubled down on what seats represent: collaboration. Notion offers unlimited blocks for individuals but imposes a 10-person visitor cap for team sharing. Trello enforces a hard 10-member collaboration ceiling; exceed it, and your entire workspace locks into read-only mode. These aren’t technical limitations—they’re monetization barriers.

Notably, the traditional tools haven’t adopted per-action pricing. They can’t. Their value isn’t in delivering computation; it’s in coordination. And coordination scales differently. A team of five people using Notion might generate 100,000 database queries a month. A team of 15 might generate 400,000. The cost doesn’t increase linearly with team size, so the business model remains per-seat. But that assumption is increasingly fragile.

When you’re under pressure to cut licenses, free users suddenly look unattractive. They consume infrastructure and support without generating revenue. Free tiers were once a reasonable growth investment. But in a market where CFOs demand immediate ROI, free users increasingly look like a liability.

This has led to what might be called the "monetization of desperation" among legacy productivity tools. Let’s look at specific examples. Evernote’s strategy is almost brutal in its clarity. The free tier has been reduced to: one device synchronization, a limit of 50 notes per month, and one notebook. This isn’t really a free product anymore; it’s a severely restricted trial designed to funnel users into the $14.99/month personal tier. The intent is obvious: the free version is so constrained that legitimate users will upgrade almost immediately.

Trello’s approach is more structural. It maintains a 10-person collaboration limit on free workspaces. Beyond that, the entire workspace enters read-only mode. This creates a sharp cliff that forces teams to upgrade. A startup might start with 8 team members on the free plan. Hire two more, and suddenly the product becomes unusable. It’s an effective, if somewhat harsh, conversion mechanism.

Notion has handled this better, and the numbers show why. Notion keeps a generous free tier: unlimited blocks, unlimited pages. But they impose a 10-person guest limit and a 5MB file size cap. The result? Notion converts around 13% of free users to paid plans—suggesting their paywall sits at the right friction point.

The common thread across all three is that free tiers are no longer customer acquisition tools. They’re customer selection tools. They’re designed to allow legitimate users to self-identify (and self-qualify) for paid plans while filtering out casual browsers.

This has led to what might be called the "monetization of desperation" among legacy productivity tools. Let’s look at specific examples.

AI Apps and the Intelligence Hierarchy

Where productivity apps are tightening their free tiers, AI applications are building entirely new value structures around cognitive capabilities. This is 2025’s most consequential pricing innovation.

The clearest example is ChatGPT’s tier structure. OpenAI doesn’t sell ChatGPT Plus ($20/month) as a better version of ChatGPT—it sells you faster access to a more capable model. You get GPT-4o by default, not GPT-4.

The implication is profound: you’re not paying for access to the software. You’re paying for access to different levels of intelligence. In this model, the product isn’t ChatGPT; the product is cognitive capacity.

Perplexity has executed a similar strategy with different mechanics. Free users get 5 "pro searches" per day—searches that include real-time web access and multi-step reasoning. Pro users ($20/month) get 300 daily pro searches. This isn’t just a quantitative difference; it’s categorical. Free users can experiment; Pro users can rely on it for actual research work.

These models diverge sharply from traditional SaaS, which sells access to data and collaboration. They’re selling access to intelligence itself, which can be rationed by cost and capability rather than by feature or seat count.

The Retention Paradox: Why AI Apps Leak Users Faster

All of these pricing strategies rest on an uncomfortable truth: users don’t stick around the way they used to. And the numbers reveal a market in transition.

AI-native applications are different. They maintain roughly 48% NRR. This gap is stunning. It means that for every dollar an AI application earned last year from existing customers, it retains only 48 cents today. The difference isn’t explained by acquisition cost or market saturation; it reflects fundamentally different retention dynamics. Users are leaving AI applications at roughly twice the rate they leave traditional tools.

There’s an important nuance here. AI applications are improving. Gross Revenue Retention (which measures cohort churn without accounting for expansion) has jumped from 27% to 40% in the first three quarters of 2025. This suggests that the "tourist" phase of AI adoption is ending; people who were experimenting are leaving, while committed users are staying. But the improvement is still fragile.

Why the churn? Several factors compound. AI applications are inherently experimental. Unlike Notion, which solves a clear collaboration problem, ChatGPT’s value proposition keeps shifting. Users adopt it for one use case, discover it doesn’t solve their actual problem, and leave.

Second, users experience what might be called the "tick-tock effect"—a psychological burden from pay-per-use models. Every query feels expensive. This friction prevents habitual use. Traditional SaaS solved this with subscriptions: you’ve already paid, so use it. AI apps sit in an uncomfortable middle where usage feels expensive but isn’t metered fairly.

Finally, mid-market AI tools live in isolation. They solve specific problems, but when those problems are solved (or cheaper alternatives emerge), there’s no switching cost. Users simply leave. Notion persists because it’s a team hub. ChatGPT persists because it’s free and habitual. Specialized tools have neither.

This dynamic is forcing a recalibration. In 2026, AI application pricing isn’t primarily about capturing value; it’s about signaling commitment and building switching costs. Charge too little, and you’re perceived as a toy. Charge too much, and you lose users before they can commit. The pricing is now part of the retention mechanics.

2026 Pricing Innovations: BYOK and Hybrid Models

A new model has emerged to solve this: Bring Your Own Key (BYOK). Users provide their own OpenAI or Anthropic API keys to the platform. The platform becomes a UI layer, a coordination tool, or an enhancement layer—while the cost of AI computation flows directly through the user’s own account.

Warp, a terminal emulator for developers, implemented this thoroughly. Users pay $20/month for the enhanced terminal experience (copilot suggestions, better search, terminal replay). When they use Warp’s AI features, those token costs bill directly to their OpenAI account, separate from the Warp subscription.

This accomplishes something elegant from the vendor’s perspective: it transfers the primary risk of AI cost unpredictability from the platform to the user. Users can experiment aggressively without worrying about a surprise bill; their costs are transparent and controlled by OpenAI’s pricing.

From the platform’s perspective, it’s almost pure margin. Collect $20/month from 10,000 users, and you’ve got predictable recurring revenue with minimal variable cost. Scale becomes a matter of infrastructure, not AI inference capacity.

But BYOK isn’t the only model gaining ground. More common is the hybrid: fixed subscription + usage credits. Users get a monthly allowance while able to purchase additional credits at set rates.

Todoist Pro exemplifies this. The subscription raised to $7/month in late 2025, and the company introduced "advanced AI features" as credit-based add-ons. Users get a small monthly allocation, but heavier AI assistance costs extra credits. By 2025, nearly 85% of SaaS companies had implemented some form of usage-based billing—most of it in hybrid form. Why the hybrid? It preserves the psychological comfort of a subscription while accommodating variance in usage. Users who barely use the AI features pay $7/month and get a comfortable experience. Heavy users can pay more but know exactly what they’re buying. It’s a compromise between predictability and variability.

These innovations exist because AI’s variable cost problem can’t vanish. But it can be distributed and priced. BYOK says: "You handle model costs; we’ll handle the experience." Hybrid billing says: "We give you a baseline; you expand from there." Both acknowledge that one-size-fits-all pricing is obsolete in an AI world.

Conclusion: Navigating the New Pricing Frontier

The 2025 SaaS market has fractured into two distinct economics. Traditional productivity applications are executing controlled retreats, tightening free tiers and collaboration limits to maintain margin. AI applications are building entirely new pricing architectures—intelligence hierarchies, outcome-based models, and shared risk structures—to cope with variable compute costs.

These aren’t competing strategies; they’re responses to different fundamental problems. Productivity SaaS has a problem of unit economics at scale: too many free users, too little margin. AI SaaS has a problem of cost unpredictability: every interaction has a different computational cost.

For product leaders in this landscape, a few principles emerge: For traditional SaaS product teams, the lesson is clear: collaboration and integration are your moats now, not features. Storage won’t hold you. Collaboration limits will. The way forward is building deeper team dependency and stickiness, then charging aggressively for it. Notion’s 13% free-to-paid conversion reflects this: they’ve built something sufficiently integrated into team workflows that people will pay to expand access.

For AI application teams, the immediate imperative is to reduce churn by building pricing that signals commitment. The NRR data is merciless: products priced below $50/month are perceived as toys; products above $250/month are perceived as strategic. The middle ground is dangerous. Consider whether you’re building a free consumer product, a committed professional tool, or a specialized enterprise application. Price accordingly, because price is now a product signal.

For enterprise buyers, 2025 is an inflection year. With nearly 80% of companies already adopting AI in at least one business unit, and CIOs demanding 20% reductions in vendor count, consolidation pressure is intense. Platforms offering integrated workflows (AI + productivity + collaboration in one suite) gain enormous leverage. Single-purpose AI tools face increasing vulnerability to replacement or integration. The irony cuts deep: in a market obsessed with AI as a cost-cutting technology, the most successful players made their pricing far more complex, less transparent, and more dependent on psychology. The free tier didn’t disappear—it transformed. The seat-based model didn’t die—it’s just inadequate now. What’s replacing it is layered, context-dependent pricing that forces every user and buyer into continuous tradeoffs about value.

This is what 2025 looks like: not disruption, but bifurcation. Two markets, two pricing logics, and an ever-widening gap between them.