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

**URL:** <https://www.artofsm.art/t/our-smaller-model-beat-opus-on-finance-tasks-alexander-shevchenko-ramp/27497>\
**Category:** Content Creators\
**Tags:** agents, machine-learning, ai-engineer\
**Created:** [4 November 2026 22:00 UTC](https://www.artofsm.art/t/our-smaller-model-beat-opus-on-finance-tasks-alexander-shevchenko-ramp/27497 "2026-11-04T22:00:34Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\
**Post date:** [4 November 2026 22:00 UTC](https://www.artofsm.art/t/our-smaller-model-beat-opus-on-finance-tasks-alexander-shevchenko-ramp/27497/1 "2026-11-04T22:00:34Z")

</div>

[![](https://www.artofsm.art/uploads/default/original/3X/c/4/c408fb76e8b2cd4bf6ee6afc09bc3dd7f30704b7.jpeg "Our Smaller Model Beat Opus on Finance Tasks - Alexander Shevchenko, Ramp") ](https://www.youtube.com/watch?v=T-42NrYjwxM)

Alexander Shevchenko from Ramp highlights the company’s applied AI research, including a specialized financial QA model outperforming frontier models, innovative KV cache compression for efficient context storage, and large-scale agentic swarms for automated security testing, all integrated into products that optimize performance and cost. He emphasizes Ramp’s commitment to advancing AI capabilities pragmatically, delivering immediate solutions while awaiting future model improvements, and invites viewers to explore their ongoing innovations on Ramp Labs’ blog.

---

<div class="post-metadata">

**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\
**Post date:** [4 November 2026 22:24 UTC](https://www.artofsm.art/t/our-smaller-model-beat-opus-on-finance-tasks-alexander-shevchenko-ramp/27497/3 "2026-11-04T22:24:44Z")

</div>

Alexander Shevchenko from Ramp discusses the company’s approach to advancing AI capabilities while waiting for frontier models to improve. He highlights Ramp’s experience over the past year, noting a 123% increase in LLM use cases and how newer frontier models have become cheaper, faster, and more accurate, enabling cost-effective deployment. Despite this, Ramp maintains an applied research function to innovate and deliver solutions promptly rather than waiting for future model releases, drawing on lessons from 2023 when building agents was more challenging due to model limitations.

Shevchenko presents three key examples of Ramp Labs’ applied research integrated into Ramp’s products. First, Ramp Sheets, launched in late 2022, is an early spreadsheet agent designed for finance professionals. To improve its performance, Ramp trained a specialized financial question-answering model based on Qwen 3.5, which outperformed the then-frontier Opus 4.6 model in accuracy and speed. This smaller, faster model works alongside larger orchestrator models to enhance user experience efficiently.

The second research focus involves KV cache compression techniques, specifically cartridges, which allow the storage of 20 to 30 times more context offline for agents. This innovation is particularly useful for recurring finance tasks like month-end close, enabling agents to start with rich, preloaded context and reducing the need for costly real-time retrieval. Cartridges also require less intensive post-training and could potentially be integrated into APIs, allowing Ramp to leverage its internal expertise immediately upon release.

The third example is the development of agentic map-reduce and agentic swarms, demonstrated by Ramp’s use of 10,000 parallel agents to perform automated penetration testing across their infrastructure. This approach identified seven high-severity vulnerabilities, including a complex Lua script injection that required understanding multiple system components. The swarm method also showed that cheaper open-weight models could be used effectively in parallel to maintain high detection rates while reducing costs.

Finally, Shevchenko highlights two Ramp products influenced by Ramp Labs’ research: token spend management, which provides detailed insights into LLM usage and costs across organizations, and Ramp Router, which optimizes agent latency and token costs through advanced routing strategies. He invites viewers to explore Ramp Labs’ blog for deeper insights into their research, including innovations in post-training transfer, self-improving monitoring, and semantic tagging, emphasizing Ramp’s commitment to saving customers time and money through continuous AI advancement.

## Useful Links

- [Qwen 3.5 Model by Prime Intellect](https://qwen.ai/blog?id=qwen3.8) — Directly substantiates the claim about the specialist financial question-answering model used by Ramp.
- [Opus 4.6 Model](https://www.anthropic.com/news/claude-opus-4-6) — Benchmark model for comparison in Ramp’s research.
