Ramp and Prime Intellect published a case study this week showing a small RL-trained model outperformed Claude Opus 4.6 by 4 percentage points on exact-match accuracy for financial spreadsheet retrieval — while running at Claude Haiku latency and a fraction of the cost. The model, a fine-tuned version of Qwen3.5-35B-A3B with roughly 3 billion active parameters, was trained in 26 hours at a compute cost of approximately $500. They called it FastAsk, a specialized subagent inside Ramp’s Sheets product that answers user questions about financial spreadsheets.
This is not a benchmark paper. FastAsk is in production, handling real finance team queries at scale. The result challenges something developers and engineering leaders often treat as settled: that the safest default for any AI task is the most capable frontier model. It isn’t always.
What FastAsk Actually Did #
According to Prime Intellect’s case study, FastAsk scored 90.4% exact-match accuracy on a held-out evaluation set, compared to 76.9% for the best frontier configuration. The base Qwen model before training sat at 64.2%, meaning reinforcement learning added 26 percentage points of accuracy in 26 hours of training. At inference time, it runs at Haiku-level latency and costs roughly $0.50 per 1,000 queries — versus $20 or more per 1,000 queries for frontier APIs. Depending on which frontier model you compare against, that is 40 to 340 times cheaper.
The training method was GRPO (Group Relative Policy Optimization), the same technique behind DeepSeek-R1 and most open-source reasoning models released in 2026. GRPO works by generating multiple candidate answers, scoring each against a reward function, then updating the model toward the better-scoring outputs. Critically, the reward function for FastAsk was simple: did the model retrieve the correct row or cell from the spreadsheet? That is a binary, automatable check — no human annotation required. Prime Intellect’s asynchronous off-policy training stack made the 26-hour wall-clock time feasible.
Related:[Alibaba’s open-code-review Beats Claude Code at 1/5 Cost]
The $500 Number Is Real But Incomplete #
The Hacker News thread hit 220 points and 64 comments within hours — and the most useful pushback landed immediately. “$500 is only compute cost,” one commenter noted. “Data generation and maintenance are the real expenses.” That is correct. The $500 covers GPU hours on Prime Intellect’s training infrastructure. It does not cover the engineering time to design the reward function, generate or validate training data, build evaluation infrastructure, or plan for retraining when the task distribution shifts.
A more complete accounting for a first-run project like this typically runs $10,000 to $50,000 when you include engineering labor — still often justified at scale, but a very different number than the headline suggests. The break-even point against frontier APIs is roughly $8,000 to $15,000 in monthly spend on a specific task. Below that threshold, paying per token is almost always cheaper once you factor in the engineering overhead.
When RL Fine-Tuning Makes Sense for Your Team #
Ramp’s case had three properties that made task-specific RL fine-tuning a good fit: the task was narrow and well-defined (spreadsheet cell retrieval), the reward signal was verifiable without human annotators (exact match), and the volume was high enough that inference cost mattered. Most tasks at most companies fail at least one of these.
However, the pattern Ramp is using — frontier models for broad reasoning and exploration, small fine-tuned subagents for high-volume narrow tasks inside a larger multi-agent system — is exactly where the industry is heading. FastAsk is not a replacement for Claude Opus 4.6 in general; it replaced Claude Opus 4.6 for one specific, high-frequency subtask inside a broader product. The distinction matters.
A rough decision framework based on industry data: below 500,000 tokens per day on a given task, frontier API wins on total cost. Between 500,000 and 5 million tokens per day, RL fine-tuning becomes worth evaluating if the task is narrow and the reward is verifiable. Above 5 million tokens per day, fine-tuning or self-hosting an open model is almost always the better option. Ramp, running a fintech product with heavy spreadsheet use, is firmly in that high-volume tier for this specific query type.
Meanwhile, enterprise AI token costs have already fallen 67% year-over-year as open-weight models narrowed the gap with closed ones. The intelligence gap between the best open-weight and best closed models is now measured in months, not years. FastAsk shows that for narrow production tasks, purpose-trained small models can leap ahead of the frontier entirely — not by being smarter in general, but by being precisely calibrated for one job.
Key Takeaways #
- Ramp’s FastAsk fine-tuned a 3B-active-parameter model to beat Claude Opus 4.6 by 4 points on accuracy, at Haiku cost and latency — all from 26 hours of RL training costing $500 in compute
- The $500 figure is compute-only; total first-run project cost including engineering typically runs $10,000 to $50,000+, with break-even against frontier APIs at roughly $8,000 to $15,000 in monthly task-specific spend
- Task-specific RL fine-tuning requires three things: a narrow, high-volume task; a verifiable reward signal (no human annotation); and sufficient token volume to justify the engineering investment
- The emerging architecture is not “replace frontier models” — it is intelligence tiering: frontier for exploration and complex reasoning, fine-tuned small models as high-volume subagents within larger agent systems