Viral AI coding assistant Cursor unwittingly unleashed a firestorm recently when it changed how it priced its product – surfacing challenges that many AI-native startups are now facing in real time.
The company had been using a request-based model: Users got a fixed number of prompts per month, regardless of how complex or compute-intensive those prompts were. As Cursor integrated more advanced models like Claude 4, a ‘request’ became an unreliable unit.
Some prompts triggered lightweight completions. Others launched multistep reasoning chains that burned through tokens and racked up real infrastructure costs. Even adjusting request counts for different models couldn’t account for the widening spread in cost and effort.
A flat-rate request no longer reflected what it cost to serve or what users got out of it. So in June, Cursor switched to a credit-based model tied more directly to compute usage. Internally, the system made more sense. But for users, it felt erratic and unclear. A handful of high-effort tasks could unexpectedly wipe out their credits, without explanation.
Cursor responded promptly with refunds and a public apology, acknowledging gaps in how the change was communicated.
Many observers framed the stumble as a communication failure, and with reason. A more phased rollout, clearer messaging by user cohort, and proactive support for those most affected could have made the transition smoother.
But it also reflected a broader issue.
One user speculated that the change may have been driven by financial pressures rather than product strategy, a disconnect that often fuels frustration when pricing shifts feel abrupt. The episode revealed how pricing has become overloaded in AI-native businesses: it’s tasked with preserving margin, driving adoption, and communicating value in products that evolve faster than users can keep up.
Of course, pricing has always had to balance growth, fiscal discipline and product evolution. But with AI, there’s no time to sequence priorities. Everything hits at once.
That’s exactly what makes a shift towards cost-based models appealing. They capture real usage and give teams an anchor to track, meter and justify pricing decisions The big question remains: Does a cost-based approach actually resolve the emerging pressures? Or does it just relocate them downstream, onto the user?
The industrywide, cost-plus reflex
Cursor isn’t alone. There’s a broader shift underway.
Most AI coding startups have been exploring how to escape the foundation model tax entirely. Cursor, Windsurf (now part of Cognition), and others are pretraining their own frontier models to reduce long-term dependence on providers like Anthropic.
But building and maintaining in-house models is expensive. For now, most companies are still paying high inference costs. That puts even more pressure on pricing to preserve margin in the short term.
According to the 2025 State of Recurring Revenue and Monetization study, 41% of companies are struggling to balance AI development costs with a coherent pricing strategy.
One by one, other AI-native companies in this space have been arriving at a similar conclusion. Lovable charges by request complexity for its agent. Some cost less than one credit, others several. When Replit introduced effort-based pricing, users quickly ran into three recurring surprises:
- Small edits in large projects triggered unexpectedly high charges due to full-context prompts.
- Long agent chats inflated future checkpoint costs in unpredictable ways.
- Prompts intended as simple questions sometimes caused the agent to modify code, incurring charges without clear user intent.
As one longtime Replit user put it on Reddit: “There’s no way to forecast what actions will cost … no control, little visibility, and no guarantee the definition of ‘work’ won’t keep shifting.”
In late 2024, Windsurf introduced a multi-credit system: one credit type for user prompts, another for behind-the-scenes actions or flow, and a flexible pool that could cover either. The structure closely mirrored how much effort each task required to run.
As the product evolved and the model’s behavior became more complex and unpredictable, those carefully defined credit ratios began to break down. Users quickly hit limits they didn’t expect or understand. By April 2025, Windsurf retired the flow credit system and reverted to a simpler, prompt-based approach. It brought back clarity, but sacrificed some of the nuance and control the original model had promised.
Each of these examples highlights the same challenge: Cost-based pricing reflects how systems operate, but not how users experience value. It may be clean from an internal perspective, but from a user’s point of view, it often feels unpredictable and hard to trust.
Some companies have defended these pricing changes by pointing out that they mostly affected a small group of power users. But user forums suggest a different story: The real issue was the nature of the tasks, such as multistep reasoning and long-context prompts, which appeared across user types, including those outside the typical ‘power’ cohort.
A common trap is equating usage with value. As investor and former operator Shelley Perry put it: “Just because they use it doesn’t mean it’s valuable.” Usage-based pricing is often mistaken for value-based pricing, when in reality, it’s simply metering activity.
The result? Revenue becomes capped by infrastructure costs, rather than aligned with what users are willing to pay. Pricing turns reactive, tracking backend expenses instead of anticipating user-perceived value.
What’s needed is a more balanced model, one that protects margin while maintaining pricing clarity, predictability, and alignment with how users actually experience value.
Yes, that model is much harder to implement in AI businesses. Models evolve fast, capabilities blur, and user outcomes are hard to trace. That’s precisely the point, though. The harder it is to see value, the more crucial it becomes to articulate it.
Why a tiered, value-aligned approach works
Cursor didn’t need to scrap the request model entirely. It needed a structure that was both legible to users and expressive of what the product actually does.
One approach would have been defining tiers that mirror the types of work users are already trying to do. For example:
- Assist → Autocomplete and quick suggestions (low cost, high frequency)
- Solve → Refactoring, debugging, and discrete code improvements (moderate cost, meaningful time savings)
- Architect → Full-file edits, multi-file coordination, or multistep reasoning (high cost, high strategic value)
These tiers translate backend complexity into something users already recognize, and they can map directly to pricing plans:
- Free Tier → 20 Assist requests/month
- Core Plan ($30–$50) → 500 Assist, 100 Solve, 25 Architect
- Pro Plan ($200+) → Unlimited Assist + expanded Solve/Architect usage
That shift matters.
Users don’t think in token burn or flow credits. They think in tasks: ‘What am I trying to get done? What kind of work is this?’ By pricing around well-understood task types, companies can capture complexity better – and set expectations without asking users to internalize system architecture.
In that sense, pricing becomes more than just a limiter or cost-recovery tool. It becomes part of the product interface, teaching users what the tool is for, setting expectations, and building trust. It doesn’t flatten or simplify AI complexity but makes it more legible.
Some core gains of value-based pricing are worth repeating here in full:
- Customer alignment: It forces you to get specific about user pain, jobs-to-be-done, and ROI, which sharpens your GTM.
- Long-term pricing power: Creates room to scale pricing with usage, outcomes or segmentation as you grow.
- Strategic signal: Shows confidence in the product’s impact, not just its tech stack.
- Future-proof fiscals: When model costs fall (and they will), your margins wouldn’t collapse with them.
The bottom line: Pricing is a tool for discovering, not just capturing, value.
Now imagine you’d been pricing for just one kind of user – only to realize the biggest impact was happening elsewhere entirely.
That was what Replit’s founder, Amjad Masad, discovered in a conversation with a public company CEO. They had assumed their product was transforming engineering workflows. But the real shift was happening in teams like product, design, and HR – functions that once waited for engineering, now building on their own. “I was surprised to hear the part about engineering teams … but it made sense – the profound impact coding agents are having on non-technical folks.”
The value they delivered – and to whom – was completely different than they expected.
A strong pricing strategy forces that kind of clarity: What kind of work is this product enabling? For whom? And what outcomes are worth paying for?
Cost-based models can’t answer those questions. They scale neatly and show fiscal discipline – but they defer the harder work of defining real value. Let’s keep doing that work.
