LegalOn halves Codex costs while maintaining development speed

On October 8, 2026, OpenAI published a customer story about LegalOn Technologies, a company that offers what it calls Professional AI to legal and other business teams around the world. After LegalOn built Codex into its development process, it cut its estimated daily costs by 65% and its costs in mature business areas by 20%, and it did so without slowing development. The story matters less for the headline percentage than for the problem it solves. Once agentic coding spreads through a company, the bill grows faster than anyone expects. Unlimited access to the strongest model drives spending up. A blanket restriction removes the productivity that justified the tool in the first place. LegalOn started in the way many early adopters do. It gave developers unlimited access to its main model, GPT-5.5 in Fast mode, and then widened its use from design to implementation to everyday work. During this phase the company learned by experiment how to divide responsibilities between people and AI. This was a sensible order of events. A company cannot write a good policy for a tool whose limits it has not yet observed. But the same source is clear about the consequence: continued unrestricted use of a high-performance model would inevitably exceed the annual budget. Cost control stopped being a finance question and became an engineering-management question. The first half of LegalOn's answer is a method for choosing models. An internal group called the AI-powered Development CoE, or AID CoE, began to write guidelines for model selection. It tested and monitored the models. Managers then passed its findings to their teams. The division of labor deserves attention. Testing and monitoring are centralized, so that the evidence is consistent. The decision itself is decentralized, so that each engineer can choose the most suitable model for the task in front of them without waiting for approval. The guidelines are not a rulebook that maps task types to models. They are a shared body of evidence that lets an individual make a good call quickly.

The choices on the menu were GPT-6 in its Astra and Luna variants, and GPT-6.1 Sol. LegalOn selected among them according to task complexity and development stage. The source material available here does not spell out which variant suits which task, and it would be a mistake to invent such a mapping. What the case does establish is the principle: the model is no longer a single company-wide default. It becomes a per-task parameter, much like choosing between a heavy tool and a light one according to the job. A simple task that does not need the most capable model should not consume the most expensive capacity, and a hard task should not be starved of capability to save a small amount of money. The second half of the answer is budget design. LegalOn aligned budgets with the growth stage of each business. This helps explain why the two reported figures differ. Mature business areas saw a 20% cost reduction, while the estimated daily cost across the effort fell by 65%. Mature products tend to have steadier, more predictable work, so there is less room to save. Businesses in earlier stages of growth need more exploration and more experimentation, so their budget logic differs. Tying spending to stage treats AI usage as an investment with an expected return, not as a flat overhead charged equally to every team. One caution is worth stating plainly: the 65% figure is described as a reduction in estimated daily costs. It is a projection, not an audited result over a full billing cycle, and readers should quote it that way.

For the wider industry, the case carries three lessons. First, competition in agentic coding is shifting from who has the strongest model to who uses models most efficiently. Model tiering and task routing are becoming basic capabilities for any organization that adopts coding agents at scale. Second, the governance that works is organizational rather than purely technical. A loop of testing, monitoring, written guidance and manager communication keeps many independent decisions consistent, and it does so without a central gate that slows people down. Third, cost and speed are not natural enemies. LegalOn reports that it lowered cost while keeping development speed, which suggests that when the selection criteria are clear, the trade-off is far smaller than the fear of it.

There is also a practical sequence that other teams can borrow. Begin with generous access so that people can discover where the tool helps. Record what you learn and assign a small group to keep testing as new models appear. Turn the findings into guidance that engineers can apply themselves. Finally, set budgets that follow the maturity of each business, so that exploration is funded where it is needed and discipline is applied where the work is stable. The specific model names will change with each release. The order of these steps is the part most likely to last.