The AI Coding Battle: From Model to Harness
Cursor’s usage-based price hikes are driving enterprises to Anthropic’s Claude Code. But smarter customers are finding ways around both, using open-source "harnesses" that significantly reduce AI coding bills. The focus is shifting from the model to the software surrounding it.
Price Shock and Countermeasures
July 29, 2026 – 3:24 pm
The most expensive part of AI coding is no longer the model itself, but the software that wraps around it. This shift has led to a fight over pricing, with Cursor’s usage-based billing model causing a stir among customers. One IT consulting firm faced a renewal bill increase from $200,000 to $1.5 million for similar usage, eventually negotiating it down to $250,000. Others, like Sanofi and Druva, also experienced significant renewals.
The Shift to Claude Code
Firms are increasingly moving their spend to Anthropic’s Claude Code, which is optimized for its own models. An IT consultancy expects its Anthropic bill to surge to at least $10 million this year, up from almost nothing in 2025. Cursor offers discounts for scale but faces limited pricing power compared to Anthropic.
Bypassing Claude Code: Open-Source Harnesses
Even Claude Code is not immune to competition. The key layer now is the "harness" software that enables code writing and editing. While Claude Code uses Anthropic’s proprietary harness, open-source harnesses exist and integrate with routers that direct simpler tasks to cheaper models. Sapiom achieved a 25% reduction in AI spending by switching to an open-source harness, while Codestrap reports a 97% cost cut after building its own platform.
"It’s not about the model anymore; it’s about the harness around the model," emphasizes Ilan Zerbib from Sapiom. Connor Deeks from Codestrap agrees: "All the value is accruing to the harness."
Implications and Future Trends
This shift represents the agentic-cost squeeze, as per-token prices continue to fall, but autonomous agents consume massive amounts of tokens, driving enterprise bills higher. Cursor faces this challenge first, but the trend suggests a future where open-source tools and innovative routing play a significant role in shaping AI coding economics.