LegalOn Cuts Codex Costs by Matching GPT-6 Models to Tasks

From unlimited top-tier access to guided model choice

LegalOn Technologies offers Professional AI globally, using AI to support legal work and other key business functions. The company integrated Codex into its development process and expanded its use across the organization through day-to-day work. As adoption grew, cost control became the main challenge. Unlimited use of high-performance models risked driving up spending, while blanket restrictions threatened the productivity gains.

Developers initially had unlimited access to the company's main model, GPT-5.5 in Fast mode. It was used for design, implementation and everyday work. Continuing that way, however, would inevitably have exceeded the annual budget. The company's AI-powered Development CoE (AID CoE) began developing guidelines for model selection. It tested and monitored models, while managers passed its findings to their teams, so each engineer could choose the most suitable model for each task.

Three models for three kinds of work

The guidelines moved the company away from using the highest-capability model for everything. Teams now start with a lightweight model and move to more capable ones as complexity increases. Internal testing made the criteria more specific:

• GPT-6 Luna, the lightest model, handles code implementation with clear requirements, everyday automations and simpler delegated tasks, mainly as a subagent.
• GPT-6.1 Sol supports standard design, data analysis and document preparation, including tasks that need shorter completion times than Luna.
• GPT-6 Astra, the most capable model, handles advanced judgment such as architecture design, complex analysis and coordinating agents as the orchestrator.

The company also introduced administrator settings with monthly usage limits for departments and individuals. AID CoE monitors usage and adjusts limits as business needs change.

Fast mode limits and budgets by business stage

Fast mode is now restricted by default, with individual requests allowed only when needed. The change raised concerns about development speed, but teams used approaches such as running tasks in parallel to keep performance up.

Budget caps now apply at the department, group and individual levels. The established LegalOn business was asked to improve cost efficiency by up to approximately 20%. New businesses in the launch phase received generous budgets to encourage active AI use. Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies, explains that these businesses prioritized business speed over cost efficiency.

According to the company, model selection, feature restrictions and tailored budgets together reduced estimated daily costs by approximately 65%, compared with GPT-5.5.

Measuring ROI per feature release

LegalOn is now asking whether faster development actually benefits customers. "It is almost a given that AI can accelerate system development and updates," says Tokitake. "What we really want to understand is whether that faster development actually translates into value for customers."

The company is building a metric that treats each feature release as a single unit, linking the customer value it delivers to the AI costs invested in it. The pipeline to calculate it is still under construction.

Next, LegalOn plans a knowledge base capturing the best model combinations for design, implementation and review, so individual successes become company-wide practice. The company has also begun emphasizing AI skills in hiring and is rethinking how work is divided between people and AI. AID CoE and the security team are working on a flexible framework that balances governance with teams' freedom to experiment.

What it means

The case, as described by OpenAI, shows one way to handle rising spend on coding agents: tier models by task, restrict premium speed modes by default, and set budgets according to each business's growth stage. LegalOn reports that this kept development speed intact while directing resources toward the businesses it wants to grow. The figures come from the company's own estimates, and its ROI metric is not yet complete.