Found first: a primary source the press has not covered yet.
A paper submitted August 15, 2026, by Ziheng Qin, Yaxin Lu, Zhangyang Atlas Wang, and Kai Wang introduces StateM, an agent-native runtime that reaches 95.3% raw accuracy on Terminal-Bench 2.1 using GPT-5.6 Sol xhigh, at an API cost of approximately $15. No model weights were modified. The performance gain is attributed entirely to harness engineering.
What the source says
StateM organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices. Running GPT-5.6 Sol xhigh inside this runtime achieves 95.3% raw accuracy across 445 trials, with all 89 Terminal-Bench 2.1 tasks solved at least once. The reported API cost is approximately $15, versus $574.68 for the reference approach. A DeepSeek-V4 Flash configuration moves from 82.7% to 88.1% under the same runtime, and GPT-5.5 xhigh goes from 83.1% to 92.1%.
Why it matters
Most benchmark comparisons treat model capability as the primary variable. StateM's results put a number on how much execution-system design moves independently: GPT-5.5 xhigh gains 9 percentage points, DeepSeek-V4 Flash gains 5.4 points, both with identical weights. The ~$15 versus $574.68 cost comparison is measured on the same 89-task benchmark, not projected from a smaller sample.