Ramp Labs applied AI engineer Alexander Shevchenko said an application-layer company can beat frontier models on cost and latency by post-training smaller open-weight models, compressing context into cached "cartridges," and running agent swarms. He said a 35-billion-parameter Qwen 3.5 model, post-trained with Prime Intellect, exceeded Opus 4.6's exact-task accuracy on financial Q&A while serving at Haiku-level latency. Ramp also ran 10,000 parallel agents to pen-test its own systems, finding seven high-severity vulnerabilities, all patched. Shevchenko said LLM use cases inside Ramp grew 123% over the past year even as per-unit token costs fell, and that cheaper tokens have made applied research more valuable because gains now sit one layer below the model, in how you post-train, cache, and orchestrate. He described the result on the AI Engineer podcast.