The Cost of AI for The End User

Companies have built incredible products but it seems that they can't figure out how to charge for them sustainably. The models we've borrowed from SaaS and consumer tech don't map cleanly to AI services where costs fluctuate wildly and value is hard to quantify.

Look at what's actually happening in the market: Subscription services race toward $0 as foundation models commoditize. The Platform tax approach (see Cursor and Droid) of adding margin to underlying AI costs feel transparent to the point of feeling extractive.

Usage-based pricing creates the kind of bill anxiety that kills product adoption. And advertising, the internet's favorite fallback, remains largely untested for AI use cases.

Based on the view from the outside, while all of these tools are sticky and people want to use them, none of these cost models have found product-market fit yet. Each reveals different tensions between what AI costs to run, what value it provides, and what users will pay for.

Subscription: The Race to Zero

The subscription model works when your costs are predictable and your differentiation is clear. AI services have neither.

Predictable revenue and user simplicity are the promise. Everyone understands $20/month. But commoditization pressure is immediate and brutal. When foundation models improve faster than application features, why pay a premium for the wrapper? Users expect more for less each month as the underlying models get better and cheaper.

The trap: you can't justify premium pricing when your differentiation is melting in real-time. Every AI subscription service is competing with slightly worse free alternatives that will be slightly better next month. Competition drives prices down faster than costs decrease, compressing margins toward unsustainability.

Platform Tax: The Transparent Middleman

Cursor and Droid take a different approach: charge a fee on top of the actual AI costs. Pass through the foundation model expenses plus a margin for the interface and tooling.

This feels honest: the costs scale with usage, users pay for what they consume, and the value stack is visible. But transparency doesn't solve the fundamental problem: costs spiral to unsustainable levels.

You start at $20, then $50, then suddenly you're paying $300 a month for what feels like a slot machine. Every interaction has variable cost. Some requests are cheap, others burn through tokens. You can't predict what your bill will be, and the unpredictability compounds with usage. The more you rely on the tool, the more anxious you become about costs. Power users, the ones who should be your best customers and advocates, become the most cost-conscious and the first to churn.

For parallel agent systems, the economics get even worse. Running multiple agents concurrently feels magical until you see the bill.

Usage-Based: The Anxiety Model

Plumb's approach was to charge per run plus actual token costs. Theoretically, this is the most fair for both the platform and the user. You pay exactly for what you use. Light users aren't subsidizing power users. Cost transparency builds trust.

But cost transparency also builds hesitation. Users prefer predictability over fairness when unpredictability means surprise bills. Billing complexity creates cognitive overhead. Every time someone considers running a workflow, they're doing mental math about what it'll cost.

This friction is poison for product adoption. The "will this be expensive?" question lives between impulse and action. Users don't estimate costs well upfront, so they either over-constrain usage or get shocked by bills. Neither outcome increases engagement.

Advertising: The Unproven Model

As I mentioned on Twitter, Amp Code seems to be the first mover in the ad-supported AI services, importing the internet's most proven business model.

Make it free for users, monetize attention at scale. You can also pay for it and have the ads removed (the YouTube model).

Free eliminates adoption friction. Advertising works across search, social, and content. The math can work at massive scale. But "massive scale" is doing heavy lifting in that sentence.

User experience degradation isn't trivial when the product is meant to help you think or work. Privacy concerns multiply when AI systems touch proprietary data or private workflows. And the fundamental question remains unanswered: do advertising economics work for AI use cases?

The model is too new to declare failure, but also too new to inspire confidence. It may require scale that most AI startups can't reach before running out of runway.

The real challenge is that most of the people they'd be advertising to are people who don't generally buy things in the software space (developers).

How Advertising Could Work for Developers

Generic display ads fail with developers, but contextual advertising aligned with their work could succeed where traditional models don't.

Sponsored API and library recommendations. When a developer asks an AI to help with authentication, show ads for Auth0 or Clerk. When they're building payments, Stripe. The AI already knows what problem you're solving—contextual relevance is built in. Developers actually want to know about tools that solve their current problem. This isn't interruption, it's discovery.

Technical job postings. Developers using AI coding tools are actively working. Companies hiring developers would pay premium rates to reach engaged engineers in context. Show relevant positions based on the tech stack in their current project. A developer writing React gets React jobs. Someone working with Rust sees Rust positions. The targeting precision is higher than LinkedIn because you're seeing actual work, not self-reported skills.

Conference and course sponsorships. Developers invest in learning. Sponsor AI responses with relevant conference tickets, online courses, or technical workshops. "This pattern you're implementing is covered in depth at React Conf" or "Egghead has a course on this exact architecture." The purchase intent is there—developers pay for education that makes them better.

Infrastructure and tooling upsells. Developers using free AI tools to build production systems need hosting, monitoring, CI/CD, and other infrastructure. Contextual ads for Vercel, Railway, or DataDog when someone's discussing deployment make sense. The AI knows they're moving to production—timing and relevance align perfectly.

How Advertising Could Work for Professionals

For non-developer professionals using AI tools, the advertising opportunities shift from technical tooling to workflow optimization and business services.

SaaS tool recommendations based on workflow patterns. When someone uses AI to organize project tasks, show Notion or Linear. When they're processing customer feedback, suggest user research tools. The AI sees their workflow needs in real-time—better targeting than any tracking pixel could provide.

Industry-specific professional services. Lawyers using AI for contract review see ads for legal research databases or e-discovery tools. Real estate agents using AI for client communication see CRM and transaction management platforms. Professionals buy tools that directly impact their revenue—contextual advertising in their workflow is less intrusive than retargeting across the web.

Executive education and certification programs. Professionals invest in credentials and skills. MBA programs, executive coaching, industry certifications—these ads work when shown to people actively solving problems that require deeper expertise. The AI can identify when someone's struggling with strategic questions versus tactical ones and serve appropriate educational options.

The model works if advertisers pay for context and intent rather than raw impressions. Developers and professionals ignore banner ads but pay attention to solutions for problems they're actively facing. The AI knows the problem—that's the entire business model's advantage over traditional advertising.

Why None of This Works Yet

Each model fails for different reasons, but the failures share a pattern: AI services are caught between three conflicting forces.

Costs are variable and often opaque. Foundation models change pricing, capabilities, and availability. What costs $X today might cost $X/2 next quarter or be unavailable next year. You can't build sustainable pricing on unstable foundations.

Value is hard to quantify. How much is a workflow run worth? What about a conversation with an AI assistant? The value depends entirely on context. The same operation is worth wildly different amounts depending on what problem it solves. Pricing that works for one use case feels absurd for another. With AI agents, every agent request feels like a tug on the proverbial slot machine arm.

User expectations are unrealistic. Everyone wants AI to be cheap or free because foundation models keep improving. The "it'll be cheaper next month" dynamic means users delay paying, delay committing, delay building workflows that depend on services that might reprice.

The companies that crack this won't just find a pricing model, they'll find a way to align costs, value, and expectations in a world where all three are moving targets.

Three Models Worth Exploring

I was curious about what's missing, so I had Claude assist me with some alternative ideas.

AI Assisted

The thinking and the writing for this part of the post was assisted by AI.

Beyond subscriptions, platform taxes, usage fees, and ads, there are pricing approaches I haven't seen tested seriously yet:

Value-Based Outcome Pricing

Charge based on outcomes rather than usage. A workflow that saves 10 hours of manual work costs more than one that saves 10 minutes, even if the AI costs are identical. This requires measuring and attributing value, which is hard—but it aligns pricing with what users actually care about. The challenge is instrumenting value capture without creating perverse incentives to inflate claims.

Dynamic Capacity Reservations

Let users pre-purchase blocks of compute capacity at different pricing tiers based on commitment. Like AWS reserved instances but for AI workflows. Users who commit to 1,000 runs/month get better economics than those who run 5 unpredictably. This smooths revenue for providers while giving users predictability. The trick is making the tiers intuitive enough that users don't need a pricing calculator.

Equity-Linked Beta Access

For products in categories without clear precedent, offer free/cheap access in exchange for equity-like upside if the user's business scales. YC for AI tooling—the provider bets on the customer's success and participates in it. This works when your tool is a genuine force multiplier for small companies who can't afford enterprise pricing. The model only makes sense when you can identify and select for high-potential users, which requires judgment most AI companies don't have yet.

Notes

• This article was partially inspired by a conversation I had with anjali_shriva in Twitter DMs that started with us talking about this thread she wrote on a token is not a stable unit of cost
• Thanks to xmok for the feedback and suggestions.
• Soren Larson responded that the analysis was good but he disagreed with the conclusion. He has a background in economics, so he's better equipped for the conclusion. He wrote about a better optimization mechanism