# SWE-Marathon ![Category](https://img.shields.io/badge/Category-Long--Horizon%20Coding-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Success-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) [![Website](https://img.shields.io/badge/Website-swe--marathon.org-0969da)](https://swe-marathon.org) > **Quick answer:** SWE-Marathon is a long-horizon, endurance-style software engineering benchmark that tests whether coding agents can sustain correct, coherent work across extended multi-step development tasks — rather than short, isolated bug fixes. Kimi K3 scores 42.0% as of July 2026. ## At a Glance **What it tests:** A coding agent's ability to maintain correctness and coherence across long-running, multi-session software engineering workflows. **Why it matters:** Many coding benchmarks test single-shot fixes. As agents are deployed for hours-long autonomous coding sessions, endurance-style benchmarks like SWE-Marathon better predict real-world reliability. **Known limitations:** As a newer benchmark with limited public documentation, task composition and scoring methodology details are not as thoroughly published as established academic benchmarks. ## What SWE-Marathon Measures SWE-Marathon evaluates coding agents on extended, multi-step software engineering tasks that require sustained focus and coherent long-horizon planning, in contrast to single-turn bug-fix benchmarks. It aims to surface degradation in agent performance that only appears over long task horizons — such as context drift, plan abandonment, or compounding errors. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Coding agent / long-horizon | | Metric | % task success | | Saturation | Low | | Created by | SWE-Marathon project | | Website | [swe-marathon.org](https://swe-marathon.org) | ## How SWE-Marathon Is Scored Agents attempt extended multi-step coding tasks, and outcomes are scored on % task success based on whether the final code state satisfies the task's verification criteria after the full multi-step workflow. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Kimi K3](/models/moonshot/kimi-k3) | 42.0% | [Kimi K3 technical report](https://www.kimi.com/blog/kimi-k3) | 2026-07 | *Score sourced from Moonshot AI's Kimi K3 announcement, July 2026.* ## SWE-Marathon on Benchgen > No Benchgen results yet — [be the first to run SWE-Marathon](/benchmarks/platform/swe-marathon). ## SWE-Marathon vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **SWE-Marathon** | Long-horizon, endurance software engineering | Low | | [DeepSWE](/benchmarks/datacurve/deepswe) | Real-world software engineering tasks | Low | | [SWE-Bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) | Professional-grade SWE agent tasks | Low | | [FrontierSWE](/benchmarks/platform/frontierswe) | Frontier-level software engineering tasks | Low | ## Run SWE-Marathon on Your Model Benchgen lets you run SWE-Marathon against your own coding agent, tracking long-horizon task success and identifying where agents drift or degrade over extended sessions. ## Frequently Asked Questions

What is SWE-Marathon?

SWE-Marathon is a benchmark testing coding agents on long-horizon, endurance-style software engineering tasks that require sustained, coherent multi-step work.

What does a good score look like on SWE-Marathon?

Kimi K3 reports 42.0% as of July 2026. Given the sustained-focus difficulty of long-horizon coding tasks, this places it among frontier-level agentic coding performance.

Who created SWE-Marathon?

SWE-Marathon is maintained as an independent benchmark project (swe-marathon.org) focused on long-horizon software engineering agent evaluation.

SWE-Marathon — Results

SWE-Marathon — Phases

SWE-Marathon