What You'll Pay for AI Agents Will Be Wildly Variable and Unpredictable

Why It Matters

Unpredictable token usage makes budgeting for AI‑driven software projects risky, threatening enterprise adoption unless providers improve cost visibility and reliability.

Key Takeaways

• Agents use up to 3,500× more tokens than simple prompts
• Same model can double token usage on identical tasks
• Token costs vary widely across models regardless of accuracy
• Hard token limits are essential to prevent budget overruns

Pulse Analysis

The University of Michigan’s recent analysis of agentic AI highlights a hidden expense that could reshape how enterprises deploy large‑language‑model assistants. While providers such as OpenAI, Google, and Anthropic publish per‑token rates, the study shows that agents—especially those built for coding tasks—consume orders of magnitude more tokens than standard chat interactions. This discrepancy stems from agents repeatedly feeding large context windows and cache reads back into the model, inflating input‑token counts far beyond the output generated. Consequently, the per‑token price list becomes a poor predictor of actual spend, leaving organizations vulnerable to surprise bills.

Beyond sheer volume, the research uncovers erratic variability even within a single model. Identical coding problems can trigger token usage that swings by a factor of two, and different models exhibit distinct cost‑efficiency profiles: Claude Sonnet‑4.5 delivers high accuracy at a premium, while Kimi‑K2 burns tokens with minimal performance gain. Crucially, agents consistently underestimate their own consumption, offering no reliable internal guardrails. For businesses, this means that traditional budgeting methods—based on average token rates—are insufficient. Companies must adopt proactive controls such as strict token caps, trimmed context windows, and minimized tool calls to curb runaway expenses.

The broader implication is a looming demand for industry‑wide standards on AI cost transparency. As enterprises scale AI‑driven development, they will pressure vendors to supply real‑time cost estimates, usage alerts, and guarantees of task completion. Without such mechanisms, the promise of agentic AI could be eclipsed by fiscal uncertainty, slowing adoption across sectors that rely on predictable ROI. Stakeholders—from CTOs to procurement teams—should therefore prioritize solutions that embed cost‑monitoring features and advocate for clearer pricing models from AI providers.