Not consistently in the coding tasks studied. A 2026 Stanford Digital Economy Lab working-paper summary reports that agentic coding runs consumed far more tokens than code chat or code reasoning, while higher token use did not reliably produce higher accuracy. That is evidence about the evaluated models and SWE-bench Verified tasks—not proof that every agent team wastes tokens or that collaboration never helps.
What the Stanford study found
The Stanford Digital Economy Lab lists the working paper, dated April 14, 2026, as an analysis of trajectories from eight frontier language models on SWE-bench Verified. It also examined how models predict the token costs of tasks. The reported figures describe that benchmark and those evaluated models, not all AI agents or workloads. Stanford Digital Economy Lab: How Do AI Agents Spend Your Money?
As an Amazon Associate I earn from qualifying purchases.
- Agentic coding used up to 1,000 times more tokens than code reasoning and code chat in the study’s comparison. Input tokens, rather than output tokens, drove overall cost.
- Repeated runs of the same task differed by as much as 30 times in total tokens. This is an upper-end run-to-run difference reported by the study, not a typical multiplier for every task.
- Model choice affected token use. On the same evaluated tasks, Kimi-K2 and Claude-Sonnet-4.5 used over 1.5 million more tokens on average than GPT-5, according to the Stanford summary.
- Cost forecasts were unreliable. The correlation between predicted and actual token costs reached up to 0.39; the summary characterizes the relationship as weak to moderate and says models systematically underestimated real usage.
Why more tokens did not reliably mean better results
The Stanford summary says accuracy often peaked at an intermediate level of token spending and then saturated as spending rose. In other words, among the evaluated runs, using more tokens was not a dependable sign that a coding task would be solved more accurately.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →This does not establish that additional computation never improves an answer. Nor does it show that every multi-agent system performs worse than a single agent: the findings concern evaluated coding tasks and models, and token use alone is not a complete measure of quality.
#1 Best Overall
Why agentic coding can consume so many tokens
The study summary identifies input tokens as the main driver of overall cost. A useful implication is that token accounting should include what an agent reads and re-reads, not just the final response it generates. But the summary does not break down every source of input usage, so it does not support assigning the observed costs to any one behavior, such as tool calls, retries, or context management.
Costs can also be hard to forecast from a single run: the reported run-to-run spread was substantial, and model estimates tracked actual token costs only weakly to moderately. A quote based on one task execution may therefore be a poor estimate of what repeated use will consume.
Rank #2
How to decide whether an agent team is worth its cost
Evaluate task success and token use together, rather than treating token volume as a quality score. For a coding workflow, record:
- Whether the task was solved, using a consistent success or accuracy measure.
- Total input and output tokens, tracked separately.
- Results across repeated runs, so cost and quality variability are visible.
- The model and agent setup used for each run.
- Predicted versus actual usage, if the system provides a cost estimate.
When comparing a single agent with a multi-agent design, hold the task and overall model budget as comparable as possible. Otherwise, any quality difference may reflect a different amount of computation rather than the team structure itself. This is a practical comparison method, not a protocol claimed by the Stanford paper.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What other collaboration research adds—and does not establish
A 2026 Nature Machine Intelligence paper titled “Capable language models can outgrow the benefits of collaboration” provides relevant context. Its available summary says it tested single-agent and multi-agent architectures across model families and agentic benchmarks, and notes that token-based communication creates latency and cost floors. The full article was not accessible for verification, so no detailed effect sizes or broader conclusions should be inferred from that summary. Nature Machine Intelligence: Capable language models can outgrow the benefits of collaboration
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
