LegalOn cuts estimated daily Codex costs by 65% with model routing

What is changing?

LegalOn Technologies stopped sending every Codex task to its most capable model and instead routed work across three models according to complexity. The company says this selection framework and budget controls cut estimated daily costs by about 65% compared with its GPT‑5.5 setup while maintaining development speed.

Under LegalOn's guidelines, GPT‑6 Luna handles code implementation with clear requirements, everyday automation, and simpler subagent work. GPT‑6.1 Sol is used for standard software design, data analysis, and document creation, while GPT‑6 Astra takes on architecture, complex analysis, and agent orchestration.

The company also restricted Fast mode by default and introduced monthly spending limits at department, group, and individual levels. Mature business units were asked to improve efficiency by roughly 20%, while newer product teams received larger budgets to preserve growth speed.

The 65% figure is LegalOn's own estimated cost calculation rather than an independent benchmark. Still, the case shows that coding-agent costs can be managed through task-aware model routing, parallel execution, and budgets aligned with a product's stage instead of through blanket usage limits alone.