The harness matters more than the model. Endor Labs tested GPT-5.5 and Anthropic Opus 4.7 in multiple harnesses the same week. GPT-5.5 scored 61.5% functional correctness in OpenAI's native Codex harness and 87.2% in Cursor's, a 25.7-point swing from the runtime alone. Opus 4.7 scored 87.2% in Claude Code and 91.1% in Cursor. Both frontier models performed better in a competitor's harness than in the one their maker ships.
The reason is cost architecture. Input tokens are 86 to 98 percent of LLM traffic on OpenRouter. Output costs 5x more per token, but input dominates the bill because volume is so much higher. The harness controls which context gets sent, and how much of it gets cached. A study across 500 long-horizon agent sessions found 41 to 80 percent cost reduction and 13 to 31 percent faster time-to-first-token through intelligent prefix caching alone. Cursor matches Claude Code technique for technique: dynamic tool fetching, priority-based prefix assembly, two-tier caching. That is why it wins on both models.
The bundled case still has real numbers behind it. Claude Code hits a 96 percent cache hit rate in production, shares system-prompt cache across users on the same version, and builds forked sub-agents at 99 percent byte-identity for 90 percent cost savings. The full piece is worth reading for how it breaks down co-design advantages and where third-party harnesses close the gap. The question it leaves open is whether model makers can defend harness quality as a moat, or whether Cursor's results already answer that.
[READ ORIGINAL →]