What happened
Endor Labs tested GPT-5.5 and Claude Opus 4.7 through different harnesses (software interfaces that shape how models process requests) and found dramatic performance swings. GPT-5.5 scored 61.5% functional correctness in OpenAI's native Codex harness but 87.2% in Cursor's harness—a 25.7-point difference from the harness alone. Claude Opus 4.7 scored 87.2% in Claude Code but 91.1% in Cursor, showing both frontier models performed better in a competitor's harness.
Why it matters
Input tokens represent 86-98% of LLM traffic on OpenRouter, making input cost the dominant expense—practitioners report a 95/5 input-to-output ratio at scale. Because harnesses control which context gets sent and can cache intelligently, they now control the cost structure more than the model itself. A study of 500 long-horizon agent sessions found that smart caching strategies delivered 41-80% cost reduction and 13-31% faster time-to-first-token, with savings scaling linearly from 500-token to 50,000-token prompts. For businesses running AI at scale, the choice of harness can rival the choice of model.
What to watch
Cursor's harness uses dynamic tool fetching, priority-based prefix assembly, and two-tier caching—techniques that match Claude Code's approach—and achieves those gains without first-party co-design. This suggests third-party harnesses may become the primary lever for cost and performance optimization, shifting power in the AI stack away from model makers toward harness builders.
Summaries like this, in your inbox every morning.
The findings challenge a long-held assumption in the AI stack: that model selection dominates performance and cost. Endor Labs' testing reveals that the harness—the software layer orchestrating how a model receives and processes queries—now exerts greater leverage. Input tokens constitute 86-98% of LLM traffic on OpenRouter, and because output costs 5× input per token but input volume overwhelms the ratio, controlling input costs is where economic value concentrates. The model itself cannot control which context is cached or sent; that power belongs to the harness. Cursor's ability to extract higher functional correctness scores from both GPT-5.5 and Claude Opus 4.7 than their native harnesses demonstrates that this leverage is not locked to first-party integration. Techniques like dynamic tool fetching, prefix assembly based on priority, and two-tier caching—all harness-level choices—can replicate or exceed the cache hit rates (96% in Claude Code's real sessions) that model makers achieved through deep co-design. For practitioners at scale, this means the harness decision may now rival or surpass the model decision in determining both cost and performance.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Google Deepmind's Gemini 4 Argon scored 97.5 percent on ARC-AGI-3 at a cost of $4,800, second only to GPT-6 As…
Cammarata reinterprets InceptionV1's cat-car neuron as a V-shape detector, not two unrelated concepts
Microsoft's Surface Laptop Ultra, using Nvidia's RTX Spark platform, starts at $2,599 (up to $5,899)
Meta's new commercial terms, effective October 30th, appear to let it crawl and train AI on content from websi…
Vercel data published Thursday shows TypeSafe AI's Jev, a lightweight decision model for choosing among predet…
Reflection AI introduced Beam, an open-weight model for agentic coding and long-horizon tasks, with 501 billio…
