--- title: "AI Pricing Under Real Usage: What Worked, What Didn't, and What's Next" description: "Early AI pricing is meeting real usage. Based on patterns across Intercom, Cursor, Salesforce, Copilot, and Jasper, this breaks down what held up, what showed limits, and where AI pricing is heading next." date: 2026-02-05 author: Kat Laszlo canonical: https://tansohq.com/blog/ai-pricing-next --- # AI Pricing Under Real Usage: What Held Up, What Didn't, and What's Next By [Kat Laszlo](https://www.linkedin.com/in/katrinalaszlo/) · February 5, 2026 Most teams didn't price AI wrong on day one. They priced it early. Now real usage is showing where those assumptions don't hold up. The data lines up with that. In [ICONIQ's latest State of AI snapshot](https://www.iconiqcapital.com/growth/reports/2026-state-of-ai-bi-annual-snapshot), only 23% of companies say they're not changing their AI pricing. 37% plan to change it in the next 12 months, and 40% aren't sure yet. That doesn't signal panic. It signals teams reacting to what they're seeing in production. When pricing does change, the direction is pretty consistent. The most common move is toward consumption-based pricing. 28% plan to go there, 21% are refining pricing based on adoption, and 15% expect to move toward outcome-based models. Pricing is drifting away from how AI was initially sold and toward how it's actually used. Margins explain why this matters. Median AI gross margins were 41% in 2024, 45% in 2025, with a projected 52% in 2026. That's improving, but it's still far from the 70–80%+ margins most SaaS teams are used to. Models that look fine at low usage start to feel very different once power users show up. Underneath all of this is a simple tension. Buyers want to pay for usage and outcomes, but they also want predictability. Those two things don't collapse neatly into a single pricing model. They push teams to try multiple approaches at once, keep what holds up, and unwind what doesn't. At this point, we've seen enough real examples to separate what worked, what didn't, and where things are heading. In every case, these approaches were reasonable starting points. Real usage is what exposed where they needed to evolve. --- ## What Worked $0.99 per resolved conversation [40% higher adoption](https://www.intercom.com/blog/pricing-ai-agents/) vs seat-based pricing Risk shifted to vendor. Customers pay when the AI solves their problem, not before. Shifted to usage-based at [$1B ARR](https://sacra.com/c/cursor/) $29.3B valuation sustained; 100x enterprise growth Costs and value stayed aligned as usage scaled. Three-model hybrid: Flex Credits ($0.10/action), Flex Agreements, per-user licenses Flexibility across use cases after initial struggles Not every customer uses AI the same way. What worked shared one trait: pricing adapted to how customers actually consumed AI. --- ## What Teams Moved Away From Flat $10/mo with unlimited access [$20-80 loss per power user](https://www.wsj.com/tech/ai/microsoft-copilot-ai-github-losses-77c71a42); 3 years to fix Usage variance became visible at scale. Heavy users consumed far more compute than light users, which stressed the model. Per-word pricing Revenue collapsed from $120M to $55-88M (~50% decline) As generation costs fell, per-word pricing became a weaker signal of value. $2 per conversation 5.3% adoption rate Undefined value metric. What's a "conversation" worth? Customers couldn't tell. What showed limits wasn't any single model. It was pricing that stayed fixed while usage patterns and costs evolved underneath it. ### Summary: AI Pricing Evolution | Company | Original Model | What Happened | Outcome | | --- | --- | --- | --- | | GitHub Copilot | Flat $10/mo | $20-80 loss per heavy user | 5 tiers + usage caps | | Cursor | Fixed credits | Hit $1B ARR, costs scaled | Usage-based pricing | | Jasper | Per-word pricing | Commoditized by ChatGPT | ~50% revenue decline | | Intercom Fin | $0.99/resolution | Pay only for success | 40% higher adoption | | Salesforce | $2/conversation | 5.3% adoption rate | 3-model hybrid | --- ## What's Next Given what worked and what didn't, these shifts are already underway. ### Cost deflation continues Sam Altman noted that GPT-4 token costs dropped 150x between early 2023 and mid-2024. "Moore's law changed the world at 2x every 18 months," he said. "This is unbelievably stronger." The January 2025 DeepSeek R1 release accelerated this. Built with a Mixture-of-Experts architecture with reported GPU costs of $5.6M, it matched GPT-4 performance at dramatically lower inference cost. If you lock in prices based on current costs, you may be overcharging in 12 months, giving competitors room to undercut you. ### "Unlimited AI" is disappearing GitHub Copilot's losses showed why. No sustainable business can offer unlimited access to costly compute. Usage caps are coming. ### Outcome metrics are becoming definable Intercom proved "resolution" works as a metric. Zendesk announced last August they're moving from $115/agent/month toward outcomes. Sierra's pitch: "pay only when we complete a task." ### Hybrid pricing is normalizing Salesforce's three-model approach isn't an edge case. Base subscription + usage/outcome components give predictable revenue while protecting margins on heavy users. --- ## If You're Building Right Now A few patterns tend to show up quickly once real usage starts. Pricing decisions get constrained by what you can actually see. Teams that can observe usage, cost, and attribution clearly have more room to experiment. Teams that can't tend to stay stuck on simpler models longer than they want to, even when they know those models won't hold up. Unlimited plans usually feel fine early on. They stop feeling fine once customers start using the product heavily. Most teams end up adding limits later, whether they plan to or not. And pricing rarely stays fixed. As models change and usage patterns settle, pricing almost always gets revisited. The companies that move fastest are the ones that expected that from the start, not the ones trying to defend a first pass indefinitely. ## Key Takeaways - ✓**What worked:** Outcome-based pricing (Intercom), usage-based at scale (Cursor), hybrid flexibility (Salesforce post-pivot) - ✕**What teams moved away from:** Unlimited flat pricing (GitHub Copilot), metrics that deflated with costs (Jasper), undefined value (Salesforce initial) - →**What's next:** Cost deflation continues (50-150x), "unlimited AI" disappears, outcome metrics become definable *[Kat Laszlo](https://www.linkedin.com/in/katrinalaszlo/) is co-founder of Tanso, flexible pricing infrastructure for SaaS and AI.* ### Want to price with confidence? Book a quick call and we'll walk you through it. More for agents: https://tansohq.com/llms.txt