Our Smaller Model Beat Opus on Finance Tasks - Alexander Shevchenko, Ramp

Alexander Shevchenko of Ramp Labs explains why an application company still runs applied research instead of waiting for the next frontier model. He walks through three projects now in production: a post-trained Qwen 3.5 model that beat Opus 4.6 on financial question answering at Haiku-level latency, KV cache "cartridges" that pack 20 to 30 times more context into an agent ahead of time for tasks like month-end close, and swarms of 10,000 parallel agents that pen-tested Ramp and found seven high-severity vulnerabilities, with cheaper open-weight models catching nearly as much. Chapters: 0:00 Intro 0:36 A bitter-lesson company 1:15 Why not just wait? 1:29 Building agents in 2023 vs 2026 3:06 Project 1: A specialist model for Ramp Sheets 5:16 Beating Opus at Haiku speed 5:58 Project 2: KV cache cartridges 7:50 Project 3: 10,000 agents pen-testing Ramp 9:30 Cheaper open-weight models in swarms 10:08 Saving time and money 10:36 Token spend management and Ramp Router 12:03 More from the Ramp Labs blog Alexander Shevchenko https://ramp.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!