---
title: "EarlyEval cuts agent evaluation costs 13–26% via early stopping"
slug: "earlyeval-cuts-agent-evaluation-costs-1326-via-early-stopping"
published: "2026-09-23"
beat: "Research"
tags: ["Research"]
creator: "Agentry Newsroom"
editor: "Susanne Sperling, Editor — Human in the Loop"
tools: ["Claude (Anthropic)", "Perplexity Sonar"]
creativeWorkStatus: "verified"
dateReviewed: "2026-09-23"
aiActArticle50: "compliant"
humanView: "https://agentry.news/research/earlyeval-cuts-agent-evaluation-costs-1326-via-early-stopping"
agentView: "https://agentry.news/agent/earlyeval-cuts-agent-evaluation-costs-1326-via-early-stopping"
---# EarlyEval cuts agent evaluation costs 13–26% via early stopping

> Researchers at Shanghai Jiao Tong University published EarlyEval on September 2, 2026, a technique that halts agent runs early by predicting task outcomes, reducing computational cost while maintainin

*Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. [AI policy](/ai-policy).*

Researchers at Shanghai Jiao Tong University published a technique for cutting the cost of evaluating AI agents without sacrificing performance, addressing a growing pain in the agent economy: the computational expense of running dozens or hundreds of agent attempts to measure progress.

**EarlyEval**, detailed in a paper posted to arXiv on September 2, 2026, uses a predictive model to identify when an agent task is likely to fail before completion, allowing evaluators to stop agent runs early [arXiv](https://arxiv.org/abs/2609.02783). The authors—Yuling Shi, Zhensu Sun, Junsen Dong, Chengcheng Wan, David Lo, and Xiaodong Gu—tested the approach on three standard agent benchmarks: SWE-bench Verified, TerminalBench, and Toolathlon.

## Cost reduction with minimal accuracy loss

The results demonstrate tangible savings for teams running large-scale agent evaluations. Across the three benchmarks, early stopping reduced agent execution steps by 13% to 26% [arXiv](https://arxiv.org/abs/2609.02783). More dramatically, the method achieved up to 44.1% fewer input tokens and 29.4% fewer output tokens consumed during evaluation runs.

The trade-off is minimal. The prediction model achieved 89% to 97% accuracy in determining whether a task would succeed or fail, and per-agent resolve rates changed by only one to two percentage points on average when tasks were terminated early [arXiv](https://arxiv.org/abs/2609.02783). That means evaluators retain nearly identical performance metrics while spending substantially fewer tokens—a direct reduction in evaluation infrastructure costs.

## Why this matters for the agent economy

As organizations scale agent deployments and iteration cycles accelerate, evaluation has become a bottleneck. Benchmarking a single model candidate across thousands of tasks in SWE-bench or TerminalBench can consume millions of tokens and hours of compute. Early outcome prediction flips that dynamic: if you can predict failure at step 3 of a 20-step task with high confidence, you stop there.

The technique is particularly valuable for organizations evaluating multiple agent candidates or versions during development. Each percentage-point improvement in token efficiency compounds across hundreds of evaluation runs. At current API costs and on-premise infrastructure expenses, a 44% reduction in input tokens represents meaningful savings per evaluation cycle.

## Implications for agent deployment pipelines

The paper does not report absolute token costs or dollar savings, but the concrete reductions—13–26% fewer steps, up to 44% fewer input tokens—are directly actionable for teams operating benchmarking infrastructure. The one-to-two percentage point variance in resolve rates is within noise for most agent development workflows, where statistical significance is measured across many runs.

EarlyEval joins a growing set of practical tools aimed at making agent iteration faster and cheaper. As the agent economy scales, reducing the cost of evaluation cycles will likely become as critical as improving model performance itself.