HarnessPro - GLM-5.2 Claude Code - SWE-bench Pro: leaderboard
HarnessPro (2026): label-free harness development, GLM-5.2 Claude Code execution setup. Official held-out evaluations; baseline GPT-5.5 high effort, GLM-5.2 max for Terminal-Bench/SWE and high for ProgramBench. HarnessPro evolves effort dynamically; Lite/Medium/Heavy are optimization cost weights, not fixed provider reasoning settings. ProgramBench uses macro-averaged within-task test pass fraction. Each method remains a distinct model-plus-harness system; cost, timing and standard deviations are not capability scores. Pass Rate (%); higher is better.
Metric: Pass Rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | HarnessPro Study GLM-5.2 Claude Code - HarnessPro-Heavy objective system | 77.3 | |
| 2 | HarnessPro Study GLM-5.2 Claude Code - RHO objective system | 75 | |
| 3 | HarnessPro Study GLM-5.2 Claude Code - HarnessPro-Medium objective system | 75 | |
| 4 | HarnessPro Study GLM-5.2 Claude Code - HarnessPro-Lite objective system | 70.7 | |
| 5 | HarnessPro Study GLM-5.2 Claude Code - Vanilla objective system | 56 |
Interactive version: theaggregate.ai/benchmark?slug=harnesspro-glm-5-2-claude-code-swe-bench-pro · How It Works · Data refreshed daily, snapshot 2026-10-21.