LegalOn Cut Estimated Codex Costs 65% by Matching Models

OpenAI published a customer story on 8 October 2026 saying LegalOn Technologies, a Tokyo-based legal-AI company, cut its estimated daily Codex costs by approximately 65% compared with GPT-5.5, while keeping development speed.

The page is OpenAI’s own case study about its customer. Its headline says LegalOn “halves” Codex costs; the body puts the cut at approximately 65% of estimated daily costs and credits three changes made together: model selection, feature restrictions and budgets set by business stage.

From Unlimited Access to a Budget Problem

LegalOn first gave developers unlimited access to its main model, GPT-5.5 in Fast mode, and widened its use to design, implementation and everyday work, the case study says.

The page then states the problem plainly: “Continuing to use high-performance models without limits, however, would inevitably exceed the annual budget.”

The company’s AI-powered Development CoE (AID CoE) began writing guidelines for model selection. It tested and monitored models, and managers passed its findings to their teams so each engineer could choose a model for each task.

Three Models, Matched to Tasks

Under the guidelines, the case study says, teams start with a lightweight model and move to more capable models as task complexity increases. The page describes the shift as moving from the highest-capability model for every task to choosing the right model for the job.

Three models split the work. GPT-6 Luna, the lightest, handles code implementation with clear requirements and runs mainly as a subagent. GPT-6.1 Sol covers standard design, data analysis and document preparation. GPT-6 Astra, the most capable, handles advanced judgment such as architecture design and coordinates agents as the orchestrator.

Administrator settings add monthly usage limits for departments and individuals, which AID CoE monitors and adjusts as business needs change, according to the page.

Fast Mode Off, Budgets by Business Stage

LegalOn also restricted Fast mode by default and allowed individual requests only when needed. The case study says the change raised concerns about development speed, and that teams ran tasks in parallel to keep performance.

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%, while new businesses in their launch phase received budgets meant to encourage active use of AI.

Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies, says in the case study that new businesses put business speed ahead of cost efficiency, with the aim of using AI extensively to raise output and drive growth.

What the 65% Measures

The body sentence carries the result: “Together, model selection, feature restrictions, and budgets tailored to each business reduced estimated daily costs by approximately 65%.”

The page’s results list phrases it more narrowly: “A 65% reduction in estimated daily costs compared with GPT-5.5 by choosing among GPT-6 and 6.1 models.” By our arithmetic, a 65% cut leaves estimated daily costs at about 35% of where they were, which is deeper than the “halves” in the headline.

The case study gives the result as a percentage of estimated costs. It states no yen or dollar amounts, token volumes or engineer counts, and it gives no measure of development speed beyond saying it was maintained.

The Next Metric: Cost per Feature Release

LegalOn now wants to know whether faster development reaches customers. Tokitake says in the case study: “It is almost a given that AI can accelerate system development and updates.”

He describes the gap the company is trying to close: “When using AI, we can track the cost of individual tasks, but the total cost of a complete piece of work—a feature release—often remains a black box.”

The company is building a pipeline that links the customer value of each feature release to the AI costs invested in it; the case study says the pipeline is not yet complete. LegalOn also plans a knowledge base of model combinations for design, implementation and review, and has begun emphasising AI skills in hiring.

The Structural Read

The starting point was unlimited access to GPT-5.5 in Fast mode. The case study says continuing that way would have exceeded the annual budget, so the change was made to fit a budget, not only to cut a price.

The ladder puts the lightest model on clear-spec code and the most capable one on architecture and agent coordination. Teams start light and move up as complexity increases, the page says, so the default became the cheapest model rather than the best one.

The budget rule is uneven by design. The established business was asked for up to approximately 20% more cost efficiency, while new businesses were given room to use AI heavily for growth, according to the case study.

Three Implications

ENGINEERING LEADERS The case study’s levers are policy, not procurement: a default lightweight model, an opt-in fast mode and monthly limits by department and individual.

FINANCE TEAMS LegalOn is building a metric that ties AI cost to each feature release, because task-level costs, Tokitake says, leave the total cost of a release a black box.

READERS OF VENDOR CASE STUDIES The 65% is an estimate of daily costs published by the model supplier, with no currency amounts or speed measure attached.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework puts it this way: “In 2026 it became a CFO problem, because the economics finally turned legible — and unforgiving.” And: “The new job of finance is to manage the ratio between the two — work produced per token burned.”

As we read it, LegalOn’s rules work on that ratio: the cheapest model is the default, a faster mode is opt-in, and the next metric the company is building prices AI by the feature release rather than by the task.

What Is Not Established

Every figure here comes from a case study published by OpenAI, LegalOn’s model supplier, and the 65% is an estimate of daily costs. We have not seen LegalOn’s own cost data; the newest releases on its Japanese news page, read on 10 October, are about other products.

The page does not say how long the period of comparison ran, or whether output per engineer changed. Nothing here says how other companies’ costs would move under the same rules.

The Bottom Line

OpenAI’s case study says LegalOn cut estimated daily Codex costs by approximately 65% compared with GPT-5.5 by matching three GPT-6-generation models to tasks, turning Fast mode off by default and capping budgets by business stage. The figure is an estimate in a vendor’s case study, and LegalOn’s next step is a metric that ties AI cost to each feature release.