---
title: 'In the News: August 26, 2026'
description: 'An independent Terminal-Bench 2.1 run shows a harness upgrade can match a newer model at a fraction of the cost, with exact transfer failing across providers.'
canonical_url: 'https://darkfactory.dev/news/2026-08-26-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-26-morning.md'
collection: news
date_published: '2026-08-26T07:45:00-04:00'
date_modified: '2026-08-26T07:45:00-04:00'
---

# In the News: August 26, 2026


A runtime upgrade raised GPT-5.5 accuracy on Terminal-Bench 2.1 from 83.1 percent to 92.1 percent, close to the newer GPT-5.6 Sol Ultra at 91.9 percent. The exact runtime profile failed to transfer across model providers, but adapting it to DeepSeek cost under $38 and improved accuracy from 82.0 to 88.1 percent.

## 1. A harness upgrade matched a model-generation jump, with costs and failures on the record

**[StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling](https://arxiv.org/abs/2608.15089)** · Ziheng Qin, Yaxin Lu, Zhangyang "Atlas" Wang, and Kai Wang, independent researchers working in their own time · arXiv, August 15, 2026

The authors built StateM, a runtime that wraps a coding agent in durable states, checked transitions, and versioned runbooks without touching model weights. Applied to GPT-5.5 on Terminal-Bench 2.1, it raised accuracy from an 83.1 percent reference to 92.1 percent, close to the newer GPT-5.6 Sol Ultra's 91.9 percent. The same runbook, frozen and unmodified, then took GPT-5.6 Sol to 95.28 percent raw accuracy in a public leaderboard submission, recorded in PR #142 on the harbor-framework repository, open and unmerged as of this writing. The authors disclose their own caveat: under stricter scoring that excludes flagged trajectories, the figure drops to 93.26 percent. Moving to a different model provider broke the exact transfer. The frozen GPT-tuned profile applied to DeepSeek-V4 Flash made results worse, from 82.7 percent to 82.0 percent. Adapting the runbook to DeepSeek cost under $38 and brought accuracy to 88.1 percent, close to the GPT-5.6 reference at a fraction of the price: about $15 in API charges for the evaluation run, against a reported $574.68 for the GPT submission.

**Why it matters:** These results isolate the harness as a variable separate from model choice and attach costs to both the successful run and the provider adaptation. They also mark a boundary: the general control structure carried across providers, but the concrete rules did not, and adapting them is what the $38 paid for.

---

## Also this cycle

- **[Vibe Coding: Practice, Performance, Productivity, and Risk, a State-of-the-Art Review](https://arxiv.org/abs/2608.20446)** · Michels, Abu Ghazaleh, Lazzari, Kassem, and Klein, KAUST and collaborators · arXiv, August 20, 2026 · A literature review of contradictory productivity claims about AI coding tools. It proposes a codebase's age as a testable variable behind the scatter and identifies a gap between audit-tier findings and the larger numbers that circulate publicly.