HarnessPro - GPT-5.5 Codex - Terminal-Bench 2: leaderboard
HarnessPro (2026): label-free harness development, GPT-5.5 Codex 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. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | HarnessPro Study GPT-5.5 Codex - HarnessPro-Heavy objective system | 80.8 | |
| 2 | HarnessPro Study GPT-5.5 Codex - RHO objective system | 76.3 | |
| 3 | HarnessPro Study GPT-5.5 Codex - RHO+Iteration objective system | 76.3 | |
| 4 | HarnessPro Study GPT-5.5 Codex - HarnessPro-Medium objective system | 76.3 | |
| 5 | HarnessPro Study GPT-5.5 Codex - ReasoningBank objective system | 74.6 | |
| 6 | HarnessPro Study GPT-5.5 Codex - Sleep-time Compute objective system | 74.6 | |
| 7 | HarnessPro Study GPT-5.5 Codex - Balance objective system | 74.6 | |
| 8 | HarnessPro Study GPT-5.5 Codex - Dynamic Cheatsheet objective system | 72.9 | |
| 9 | HarnessPro Study GPT-5.5 Codex - HarnessPro-Lite objective system | 72.9 | |
| 10 | HarnessPro Study GPT-5.5 Codex - Vanilla objective system | 69.5 |
Interactive version: theaggregate.ai/benchmark?slug=harnesspro-gpt-5-5-codex-terminal-bench-2 · How It Works · Data refreshed daily, snapshot 2026-10-21.