Cut Codex spend without slowing down: LegalOn starts on the light model, steps up only when the task needs it

#What changed

LegalOn Technologies first gave its developers unlimited use of its top model in Fast mode. Spending was on track to blow past the annual budget, so an internal AI group wrote model-selection guidelines. Engineers now start on the lightest model and move up only when the task calls for it. OpenAI's case study puts the result at about 65% lower estimated daily costs with development speed held steady, and about 20% better cost efficiency in the company's mature business.

#Who is affected

Anyone whose Codex bill or usage limit grows faster than their output, especially teams that default every task to the most capable model or keep Fast mode on all the time.

#What to do now

• Match the model to the job. LegalOn uses GPT-6 Luna for clearly specified implementation and as a subagent worker, GPT-6.1 Sol for routine design, analysis, and documents, and GPT-6 Astra for architecture, complex analysis, and orchestrating other agents.
• Start light and escalate. Begin a task on the cheaper model and switch up only when it struggles, instead of starting at the top.
• Make Fast mode opt-in. LegalOn turned it off by default and allows it on request; teams kept their pace by running tasks in parallel.
• Set budgets per department, group, and person with admin usage limits, and give new projects more room than mature ones.

#What is not confirmed

The 65% and 20% figures are LegalOn's own estimates as reported by OpenAI, measured against its earlier GPT-5.5 Fast setup. DotsFeed has not independently measured them, and your savings will depend on your task mix.