jevagentharness/jevagentharness is a local OpenAI-compatible proxy that has a small decision model classify each step of a coding agent as 'ordinary' or 'hard' and then assign a model. Ordinary steps go to a fast model, hard ones to a strong model. According to the README, Python 3.10 or higher is required and the project is at alpha stage.
The README publishes the results of running the 74 held tasks of Terminal-Bench 2.0 once per harness. For mini-swe-agent, the success rate dropped from 66.2% to 63.5%, a 2.7 point decrease, and the 95% interval was -14.9 to +8.1. Cost per task fell from $0.91 to $0.33, a 64% reduction, and time fell from 13.7 minutes to 7.9 minutes, a 43% reduction. pi and DeepSeek Harness cut cost by 28~31% and time by 4~15% with no drop in success rate. However, the share of steps that remained on the strong model was 45~55%.
The size of the savings depended on the price difference of the fast model. On Flash, which is 2.7x cheaper per token, savings were only 7~13%, and on Flash-Lite, which is 4~7x cheaper, they were 28~64%. In Terminus-2, 53 of the 74 baseline tasks ended in timeouts, so the author states that no conclusion could be drawn. The protocol registered before the run, the raw per-task data, arXiv 2610.21412, the models and the data are all public. That said, these are the author's own measurements from an alpha project, and performance cannot be judged from the Terminus-2 results.