LegalOn Technologies says it reduced estimated daily Codex costs by roughly 65% without slowing development by selectively routing tasks to different models and tightening budget controls. The company matched GPT‑6 Astra, GPT‑6.1 Sol and GPT‑6 Luna to task complexity, limited Fast mode by default, and applied monthly usage caps at department and individual levels.
Model matching and governance
The company integrated Codex into its development workflow and then the AI-powered Development CoE (AID CoE) created guidelines for model selection. Engineers began using lighter models for routine work and escalating to more capable models as complexity rose. According to the company, GPT‑6 Luna handles code implementation and simple automations; GPT‑6.1 Sol supports standard design, data analysis and document preparation; and GPT‑6 Astra is reserved for complex architecture, advanced judgment and orchestration roles.
To enforce the policy, administrators set monthly usage limits and defaulted Fast mode off, permitting higher-cost requests only when justified. AID CoE monitors consumption and adjusts limits to reflect business priorities, the company states.
Budgeting, performance and measuring ROI
LegalOn combined model selection with budget caps that increased efficiency in mature businesses by about 20% while allocating larger budgets to new launches to prioritize speed. The company reports an overall 65% reduction in estimated daily costs compared with using GPT‑5.5, achieved by choosing among the GPT‑6 and GPT‑6.1 family for different tasks.
Leaders at LegalOn also raised the question of whether faster development creates customer value. Yuta Tokitake, Senior Engineering Manager, said the company is building a pipeline to measure ROI per feature release, linking the customer value of each release to the AI costs invested. The goal, he said, is to make return on investment visible and accurate.
The company is also compiling a knowledge base to capture best practices for model combinations and is emphasizing AI skills in hiring. AID CoE and the security team continue to adjust controls to balance experimention, cost and risk while directing resources toward prioritized businesses.
Original source: OpenAI News