Release2026-09-23
Cursor cut token costs for its agent harness by 7% without hurting quality. It trimmed its system prompt by 66% as models improved. It now loads tool definitions only when needed. It added explicit cache breakpoints for GPT-5.6 to reuse more context. It reduced line-number tokens in file reads by 1.6%. It offloaded built-in tools to dynamic context, cutting static tokens by 60%. It adjusted subagent usage to avoid redundant work. It also applied these learnings to Grok Bot.
What changed
Cursor's agent harness included full tool definitions and verbose system prompts in every request.
What it unlocks
Running longer agent tasks at lower token cost without quality loss.
- 7% token cost reduction
- 66% system prompt trimmed
- 46.9% token reduction for MCP tools
- 60% static-context description tokens cut
Sources