Ramp's Fast Ask beats Claude Opus in Sheets

Hot take: this is a strong proof that narrow, workflow-specific RL can beat frontier generalists on high-value enterprise tasks without needing a bigger base model.

• –The win is not “general intelligence”; it is targeted retrieval quality on one ugly but important spreadsheet workflow.
• –The training setup matters: custom environment, tool use, and constrained workbook navigation are what made the improvement measurable.
• –The reported gains are commercially meaningful because the model is faster and cheaper, not just more accurate.
• –The strongest signal here is product strategy: build a specialist subagent for a repeated bottleneck instead of routing everything through a general-purpose model.