Cost per task

• Pricing or margin analysis for AI features.
• Evaluating prompt/model changes that affect token usage.
• Deciding whether to add steps like reflection or reranking.

Cost per task anchors unit economics. PMs must ensure quality gains justify cost increases. Lowering cost can unlock scale or freemium tiers; raising cost may require paywalls or usage caps. Cost also influences model/provider choices.

Log tokens and tool spend per task, including retries. Track by intent and customer segment. In 2026, set guardrails that block deploys when cost spikes without quality lift, and auto-route to cheaper models for low-risk intents.

Switching simple intents to a smaller model cut cost per task from $0.12 to $0.05 while TSR stayed at 79%, enabling a free tier without margin loss.

• Ignoring tool/API fees when calculating cost.
• Averaging across intents instead of focusing on the expensive few.
• Deploying quality improvements without checking ROI on cost/latency.