Not Diamond

building ¬◇, intelligent model routing for coding agents

Not Diamond Code is a model router built specifically for long-horizon coding agents, aimed at engineering teams whose bills have ballooned because their agents default to the most expensive frontier model for every step. The product sits between the harness and the underlying models, predicting which model and how much reasoning effort each turn actually needs, then routing accordingly. It plugs into any gateway or harness, including Claude Code, and the company claims cost reductions of 20 to 65 percent without giving up output quality. The timing tracks with a real pain point around Claude Code and similar agent harnesses, where locking in one provider gets expensive fast and switching per task is fiddly to do by hand. Not Diamond's pitch to developers is that its router intelligently predicts which model to use for each input to reduce cost while maintaining accuracy, and integrates with existing harnesses and gateways to avoid vendor lock in. The company already lists Hugging Face, Dropbox, IBM, DoorDash, OpenRouter, and Rootly among the teams using it, and cites a 39 percent average accuracy lift on Rootly's SRE benchmarks. Not Diamond was founded by Tomás Hernando Kofman (CEO), Jeffrey Akiki, and Tung Tze-yang (CTO), and raised roughly US$2.3 million to build out the router. Kofman is posting the launch, framing Not Diamond Code as the coding-agent-specific extension of the broader routing work the team has been shipping for general LLM workloads.

Tags

Comments (12)