LegalOn Technologies reduced estimated daily Codex costs by about 65% by selecting models based on task complexity, restricting Fast mode, and allocating budgets according to business stage. According to a case study published by OpenAI on October 8, 2026, the company says it maintained its development pace. The next step is to link AI costs to the value of each feature release.
Not using the most powerful model for everything
Initially, LegalOn gave developers unlimited access to GPT-5.5 in Fast mode. This approach helped the team experiment with AI in design and coding, but created a risk of exceeding the annual budget.
Rather than imposing blanket restrictions, the department responsible for AI-assisted development created model selection guidelines. Teams start with a lightweight model, then move to a more powerful one when tasks become complex.
In the division of work described by LegalOn, GPT-6 Luna handles code implementation with clear requirements; GPT-6.1 Sol supports routine design, data analysis, and document drafting; and GPT-6 Astra handles architectural design, complex analysis, and agent orchestration.
This is an internal division of work, not evidence that each model is always the best choice for every business. Evaluating Astra before deployment still needs to be based on specific tasks and quality criteria.
Different budgets for different business stages
LegalOn also restricts Fast mode by default, allowing it only when there is a specific need. The company says teams maintain productivity through approaches such as running multiple tasks in parallel.
Usage limits are set at the department, team, and individual levels, then adjusted according to need. Established business areas were asked to improve cost efficiency by about 20%, while new areas received larger budgets to prioritize development speed.
The reduction of about 65% therefore resulted from a combination of changes, not just switching models. This case differs from Asana's optimization of browser agent history and caching, where the focus was on how the workflow uses context.
From AI bills to feature value
LegalOn is building a system to measure AI costs for each feature release, rather than tracking only individual tasks. The goal is to compare the investment with the value customers receive.
Notably, this measurement system is not yet complete. The source does not provide feature-level return-on-investment results, so it is not yet possible to conclude that the savings have translated into corresponding business benefits.
The 65% figure is also described as a reduction in estimated daily costs compared with the previous use of GPT-5.5. The publication does not fully detail the measurement period or provide an independent set of metrics for development speed and quality. These are company-reported results in a vendor publication, not savings that can be assumed to apply elsewhere.
For organizations that are expanding employee-built AI applications, LegalOn's case shows that governance goes beyond usage limits. The open question is how much each AI expenditure contributes to the product that is actually released.