We ran Terminal-Bench 4.0 11% cheaper and 12% faster than the frontier labs (www.omnara.com)

Omnara has announced impressive results with its open-source harness, achieving 11% lower costs and 12% faster performance than traditional labs while running the Terminal-Bench 4.0 benchmark. This benchmark was performed with two advanced models, Claude Opus 5.5 and GPT 6.1 Sol, where Omnara's system not only matched pass rates but also outperformed the proprietary harnesses—Claude Code and Codex—in efficiency. The Omnara platform operates in the cloud, allowing for enhanced scalability and process management, streamlining tasks through a unique tool architecture that minimizes idle time and reduces the number of necessary API calls. This achievement is particularly significant for the AI/ML community as it demonstrates that high-performance machine learning infrastructure can be developed outside major labs, potentially democratizing access to effective AI deployment tools. Omnara’s results indicate that strategic optimizations, such as fewer LLM API calls and deferred context management, can generate considerable cost benefits—saving around 5% compared to Claude Code and offering substantial token efficiency with GPT 6.1 Sol as well. These advancements could pave the way for wider accessibility and better resource utilization in AI deployments, making it an appealing option for researchers and developers alike.