LegalOn cut estimated Codex costs by 65%. Task routing did the work

OpenAI says in a customer case study that LegalOn reduced estimated daily Codex costs by about 65% compared with using GPT-5.5. The company also says it maintained development speed and quality. It attributes the result to model selection, feature restrictions and budgets tailored to different parts of the business.

Three models split the work by difficulty

LegalOn stopped sending every task through the most capable setup. GPT-6 Luna handles implementation from clear requirements, routine automation and simpler analysis. GPT-6.1 Sol covers standard design, analysis and document creation. GPT-6 Astra takes architecture, harder judgment calls and orchestration of other agents.

The company also disabled Fast mode by default and makes it available on request. Monthly limits apply to departments and individuals. The established business was asked to improve cost efficiency by up to 20%, while newer business units received looser budgets to encourage experimentation.

Coding-agent spend is becoming an operations problem

The useful lesson is not the Astra, Sol and Luna branding. Teams need to separate cheap routine changes from work where a bad result increases review, testing or production costs. A low-priced run is poor value if it creates another repair cycle.

LegalOn therefore presents an internal operating model rather than a universal discount. Someone must continually test models, write routing guidance and adjust limits. API savings can disappear into human labor when a company measures tokens but ignores accepted changes.

The 65% figure still rests on vendor methodology

OpenAI does not disclose absolute spend, developer count, task volume or the observation period. It also provides no metric demonstrating maintained speed and quality. The comparison is an estimate of daily costs against GPT-5.5, not an audited bill across equivalent periods.

Without review time, defect returns, test success and accepted-change data, readers cannot tell whether some cost merely moved from the model to developers. The case study describes a useful practice, but 65% is not a portable benchmark.

Accepted changes will reveal whether the cheaper route saves money

The next useful signal is cost per accepted change or shipped feature, a measure LegalOn says it is still developing. Daily spend needs to sit beside cycle time, review load, rework and post-deployment incidents.

If lighter models hold those measures steady and the strongest model appears only when needed, routing becomes a repeatable operating tool. If manual checking grows, the saving will be little more than a neatly arranged token bill.

Lilith's verdict

LegalOn put the strongest model in the control room and lighter models on routine shifts. The saving becomes real only if developers do not find a pile of repairs waiting at the end of the day.