StateM: 95.3% on Terminal-Bench or a $15 Frontier Run via Harness

The paper reports that a runtime layer alone pushes a frontier model to 95.3% raw accuracy on Terminal-Bench 2.1, and separately gets a cheaper model to 88.1% for $15 in API costs against a $574.68 GPT reference run. No new weights, no fine-tuning.

The setup is what the authors of the arXiv preprint call StateM: a runtime organized around, in their words, "durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices." Their diagnosis of long-horizon agent failure is blunt: "Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedures, or stop prematurely."

The 95.3% number comes from GPT-5.6 Sol xhigh across 445 trials, with the system succeeding on all 89 tasks at least once. GPT-5.5 xhigh reached 92.1% versus an 83.1% reference baseline. GPT-5.6 Luna moved from 76.7% to 85.4%. On the cheap side, DeepSeek-V4 Flash climbed from 82.7% to 88.1%, with total DeepSeek expenditure of $52.22.

On BusinessBench, family-specific runbooks yielded held-out gains of 0.55 to 10.04 points depending on how closely a task's structure matched the runbook.