|
Cursor, the pioneering AI coding app, is growing pretty well ahead of SpaceX’s planned $60 billion purchase. But the company, which provides access to different AI models for coding work, last year shifted to charging more customers based on usage rather than flat rate subscriptions, and the move is now causing some heated contract renewal discussions. That seems to be pushing some customers to shift to Claude Code and other AI coding apps. We talked with an executive at an IT consulting firm who said they paid Cursor around $200,000 for 800 licenses in the 12 months ending in May of this year. When the contract came up for renewal, Cursor tried to charge the IT firm around $1.5 million for essentially the same amount of AI usage, according to this person and documents viewed by The Information. The IT firm “played hard ball” and negotiated the new yearlong contract down to $250,000, the person said. Cursor’s acceptance of that offer shows how it may not have the level of pricing power enjoyed by other AI firms, most obviously Anthropic, at least for now. Indeed, the IT firm’s executive expects the firm’s Anthropic bill to soar to at least $10 million this year, driven by its usage of Anthropic models through Claude Code, from a “negligible” amount in 2025. The IT firm believed it could get more value from using Claude models through Claude Code than through Cursor. Whether that’s right is debatable. Cursor recently released a model router that helps customers swap between cheap and expensive models, depending on the complexity of the coding task, which Cursor says will save customers significant sums. Right now, though, some Cursor customers are complaining about big price increases. They include data-security software provider Druva and Sanofi, a biopharmaceutical firm. Sanofi Chief Digital Officer Emmanuel Frenehard told me that Cursor recently proposed a contract renewal that would cost five times what the company paid the startup last year. “I may not renew with Cursor,” said Frenehard, who has not yet decided. “I get a better deal with Claude Code.” That’s because Claude Code is optimized to use Anthropic’s models, though engineers can also use Claude models for coding tasks through Cursor. The customers I spoke with said they didn’t want to pay to use Anthropic models through Cursor without a significant cost benefit. Despite the hand-wringing, data from model routing services provider Weave show its customers have actually seen their Cursor costs rise to a lesser degree than Claude Code over the last six months. Claude Code also has shifted toward usage-based pricing, so perhaps it has more leverage over customers than Cursor does. Suffice it to say there’s room for more than one winner in the AI coding market! Anthropic Customers Explore Ways to Lower Claude Code Costs We’ve written extensively about how companies are finding ways to lower the cost of running AI agents and chatbots. But until recently, most companies we talked to were gung-ho about using the most advanced Anthropic models for coding, as I reported today. Now, some businesses are starting to use open source harnesses—the software that helps AI models write and edit code—rather than Claude Code, which uses Anthropic’s proprietary harness. Doing so helps companies save money because the open source harnesses can easily hook up to model routers that push simpler workloads to open source or smaller proprietary models that are less expensive to run, compared to Anthropic’s. In recent weeks, Sapiom, a startup that sells software to manage AI agents for financial transactions, used new harnesses and reduced its Claude Code usage, CEO Ilan Zerbib said. He said he expects the move to cut his engineering team’s AI spending 25%. That team now uses OpenCode, an open source harness, and also developed its own harness using software from Pi, a startup. In some cases, engineers still use Claude Code and connect it to open source AI models, though that’s not recommended by Anthropic because its application is optimized for Anthropic models. “It‘s not about the model anymore; it’s about the harness around the model,” Zerbib said. Notably, Anthropic has invested in his startup. Connor Deeks, the CEO of AI coding startup and boutique tech consultancy Codestrap, said nearly the same thing after he began using an internally developed AI coding platform to build custom AI software for clients and compared it to Claude Code costs. He says AI coding costs among the customers plummeted 97% since then, in part because the coding platform includes a harness that routes many tasks to small models from Google and OpenAI. “All the value is accruing to the harness…it has nothing to do with the model,” he told me. AI Adoption Gap Workers are still picky about when and how to use AI that’s sold by traditional software firms, according to our interpretation of a new report from Banyan Software, which owns and operates more than 100 enterprise software firms. Roughly half of the 260 software firm executives Banyan surveyed this year said fewer than 25% of their customers are even using AI features the software firms launched this year. (The software firms have less than $50 million in revenue.) The data show there’s a lag in companies creating AI features and when customers actually use them. While companies have lapped up coding tools from Anthropic, they’re not ready to use every AI feature under the sun.
| 
