Cursor's 7.5x renewal ask is pushing customers to the harness layer
AI coding bills stopped being model bills this year. The enterprise contracts now moving tell you where the money is actually going.
One IT consultancy paid Cursor about $200,000 for 800 licences in the year to May. At renewal, Cursor asked for roughly $1.5 million for the same usage — a 7.5x jump. The firm negotiated the ask down to $250,000, a quarter of the original price but still more than it paid last year. That number comes from The Information's reporting on Cursor's move to usage-based pricing, and it is not an outlier: Sanofi and the security vendor Druva both faced renewals about five times higher than the previous year. Sanofi's chief digital officer Emmanuel Frenehard gave the short version of how enterprises are responding — "I may not renew with Cursor. I get a better deal with Claude Code."
The near-term winner of that reaction is Anthropic. Claude Code is tuned for Anthropic's own models, so firms shifting spend there get a coherent stack; the same IT consultancy now expects its Anthropic bill to reach at least $10 million this year, against nearly nothing in 2025. But the reporting undercuts the obvious conclusion that one premium vendor simply replaces another. Data from the routing firm Weave shows Cursor bills actually rose less than Claude Code bills over the past six months, and Cursor now ships its own router that sends simpler tasks to cheaper models. Per-token prices keep falling while agent bills keep climbing, because autonomous coding burns tokens in volumes that price cuts cannot offset — and that arithmetic applies to any premium tool sitting on a commodity model layer.
The real movement is one layer down: Sapiom, a startup Anthropic has backed, swapped Anthropic's own Claude Code for the open-source OpenCode harness and expects to cut its AI spending by about 25 percent. Codestrap, a coding consultancy, says customer costs fell about 97 percent after it built its own platform and routed most work to small Google and OpenAI models. "It's not about the model anymore," Sapiom's Ilan Zerbib told The Information. "It's about the harness around the model." Codestrap's Connor Deeks was blunter: "All the value is accruing to the harness. It has nothing to do with the model."
Even Anthropic's own tool is being pried open. The same reporting found developers using the Claude Code harness to drive cheaper non-Anthropic models — including GPT-5.6 Sol — through proxies and routers like OpenRouter. That runs against Anthropic's design: OpenRouter's own documentation notes that its native Claude Code integration is guaranteed only against Anthropic's first-party API, and that non-Anthropic models are not supported through that endpoint, which is exactly why community proxies exist. The harness is the last defensible part of the stack, and it is being commoditized in public.
We mapped the same economics from the supply side this morning — The closed-model premium buys 4.4 months — then it's lock-in — and the routing fight showed up earlier this month in GitHub's HydraFusion routes between models to match Opus 5 at 67% less cost. The new detail is that the buyer is now doing the unbundling, contract by contract, instead of waiting for a vendor to offer it.
What to watch: whether Cursor's renegotiated renewals hold at 1.25x next cycle, whether Anthropic starts metering Claude Code per task rather than per seat, and whether any enterprise publishes what its custom harness costs to maintain — because a 97 percent saving that requires a platform team is not a 97 percent saving.
If your AI coding renewal came in 5x higher, would you pay up, switch vendors, or build the harness yourself? Tell us in the comments.
