How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks

Agentic coding tasks are uniquely expensive, consuming orders of magnitude more tokens than code reasoning and code chat tasks, and models vary substantially in token efficiency across different frontier LLMs. This systematic study analyzes token consumption patterns in agentic coding tasks using trajectories from eight frontier LLMs on SWE-bench Verified to address where agents spend money, which models are most efficient, and whether agents can predict their own costs before execution.

Key Findings on Token Consumption Patterns

The analysis reveals several critical insights into the economics of AI agents. First, agentic tasks are uniquely expensive, consuming 1000× more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost. Second, token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30× in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs. Third, models show significant variation in efficiency: on the same tasks, Kimi-K2 and Claude Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5. Finally, a fundamental gap exists between human perception and computational effort: task difficulty rated by human experts only weakly aligns with actual token costs.

Token Usage Dynamics Across Tasks and Runs

The study systematically compared agentic coding against code reasoning and code chat tasks, showing that agentic workflows accumulate the information from different sources and the same context gets fed into the models repeatedly, resulting in a dramatically higher input/output ratio than the other two task types. Furthermore, token usage exhibits high variance both across problems and repeated runs of the same problem. The most expensive problem costs ∼7 million more tokens than the cheapest, and high-token-cost problems exhibit larger cross-run variance, suggesting that agent behavior becomes increasingly unstable on more complex tasks. This variability makes upfront cost prediction fundamentally difficult.

Accuracy vs. Cost Trade-offs

The research demonstrates a non-monotonic relationship between token expenditure and performance. When working on the same problem, agent performance peaks at the intermediate-cost run and then saturates with higher costs. This is consistent with findings that additional reasoning steps or longer chains of thought do not necessarily improve accuracy, suggesting that increased computation may reflect unproductive exploration rather than deeper reasoning. Moreover, analyzing fine-grained actions shows that high-cost runs are associated with repeated view and edits for the same file, indicating that these expensive runs are often dominated by redundancy rather than productive progress.

Predicting Agent Token Consumption Before Execution

The paper formalized the task of predicting token consumption before execution, focusing on self-prediction where the agent estimates its own usage. The results indicate that agents can capture coarse trends in token consumption, but achieve only weak-to-moderate correlation with real usage across models. Specifically, output-token usage is easier to predict than input-token usage, reflecting the uncertainty introduced by context construction. Although accurate instance-level prediction remains challenging, self-prediction provides a useful coarse-grained signal of relative cost that can support early budget alerts.

Model Token Efficiency and Prediction Performance

Models vary substantially in their token efficiency across tasks, with efficiency differences stemming from model-specific behavior rather than intrinsic task difficulty. For instance, GPT-5 and GPT-5.2 achieve strong accuracy at low cost, while Kimi-K2 remains an outlier with both the highest cost and the lowest accuracy. In terms of prediction capability, Agents systematically underestimate both input and output token usage, a bias that persists even when no in-context example is provided. While self-prediction is challenging, it offers a measurable signal for cost transparency without requiring additional infrastructure.

Implications for Pricing and Future Research

The findings highlight the tension between variable consumption and the need for predictable pricing models. Because agent trajectories are inherently stochastic, purely upfront pricing remains difficult, suggesting that consumption-based pricing will likely stay the most practical option until pre-execution estimation becomes more reliable. The study suggests that agents themselves can potentially serve as useful predictors of their own cost, at least at the coarse-grained level of identifying high-cost tasks, which is sufficient for providers to issue early warnings. This work provides a foundation for future research aimed at building more controllable and transparent agent pricing schemes.

Conclusion

Self-prediction captures coarse-grained trends in token usage but remains noisy at the instance level. Overall, the study confirms that predicting token usage before execution is a genuinely difficult task for current models, yet it offers concrete steps toward more predictable and user-aligned agent pricing by identifying high-cost tasks.

**(Note: The provided text does not contain a section explicitly titled The gist or a single sentence summary under that heading.

Here are specific improvements to AI systems based on the findings of this research, along with what those improved systems could achieve:

1. A mechanism for Pre-Execution Cost Bidding for Agentic Tasks: The system should not execute a long-horizon agent task (like complex coding) without first running a self-prediction sub-routine that estimates the total expected token cost and confidence level.

2. Dynamic Model Selection Based on Predicted Budget: Agents should be able to query multiple frontier models and select the one that offers the best trade-off between accuracy and predicted cost, potentially switching to a more token-efficient model if the task complexity is high or if the predicted budget is exceeded.

3. Adaptive Stopping Criteria Guided by Cost Thresholds: The agent's internal planning mechanism should be augmented to monitor accumulated tokens against a user-defined or pre-set budget threshold (derived from the initial prediction). If the cost approaches this threshold, the agent should automatically trigger a Plan to Stop phase, prioritizing summary generation over further exploratory actions, mitigating runaway token consumption.

4. Cost-Aware Task Decomposition: Before starting complex tasks, the agent should decompose them into sub-tasks and assign an estimated token budget to each phase (e.g., Exploration vs. Debugging). This allows for more granular cost tracking and prevents a single high-cost exploratory step from bankrupting the entire task's budget.

5. Transparency Layer for User Trust: Implement a user-facing interface that displays the predicted token cost and confidence score alongside the task description before execution begins, allowing users to make informed decisions about committing to an expensive run.

This improved AI system can achieve:

1. Reduced Financial Risk for Users: By providing upfront cost estimates, users can avoid unexpected high bills from long-running or failed agent tasks.

2. Optimized Resource Allocation: The system ensures that computational resources are used efficiently by favoring models with better token efficiency for the perceived complexity of the task, leading to lower operational costs for developers deploying these agents at scale.

3. More Robust and Reliable Agents: By incorporating cost constraints into the planning loop, agents become inherently more cautious about redundant or unproductive exploration, leading to faster convergence on solutions and higher success rates for a given budget.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.