Matching Codex Models to Tasks Cut Daily Costs by 65%
LegalOn assigned different models to simple coding, documentation, and complex design work, then set spending caps by department. OpenAI says daily costs fell sharply while development speed and quality held steady.
On October 8 local time, OpenAI published a case study on how LegalOn Technologies runs Codex. The company now assigns developers different AI models depending on the kind of work at hand, and it has reworked how usage budgets are allocated across departments and individuals. According to OpenAI, the changes cut daily costs by roughly 65% compared with using GPT-5.5, with no loss in development speed or quality.
Which model for which job
The type of task is the deciding factor. Coding with clear requirements, simple jobs, and automation go to GPT-6 Luna. General design, analysis, and documentation run on GPT-6.1 Sol. Complex analysis and software architecture design fall to GPT-6 Astra. Astra also handles the job of coordinating multiple agents, the AI systems that carry out work on their own.
None of these distinctions existed at first. Developers had unlimited access to GPT-5.5’s Fast mode. The company later introduced a pattern of starting work on a lighter model and switching to a higher-tier one as the task grew more complex. Fast mode became blocked by default, available only by individual request.
How the budgets were divided
Usage limits are managed at two levels: the organization and the individual. Spending caps are set for departments, groups, and individuals, with monthly usage limits applied to departments and individuals as well. Limits are adjusted as needs change. In effect, the model selection criteria come paired with a throttle on consumption.
Allocation also varied by business stage. Established business units were asked to improve cost efficiency by up to roughly 20%. Newly launched business lines, by contrast, received relatively generous budgets to encourage heavier AI use. Within the same company, different standards apply to units expected to cut costs and units expected to expand AI usage.
Where the 65% figure comes from
The roughly 65% reduction reflects the combined effect of model selection, Fast mode restrictions, and budget allocation. OpenAI described it as an estimate of daily costs compared with using GPT-5.5. The company also said development speed and quality were maintained, but it did not release benchmark scores to back that up.
For organizations in South Korea using AI in software development, the most instructive part may be how finely the criteria were split. Simple coding, documentation, and complex design each got their own model, and usage limits were tuned to the business stage of each unit.
LegalOn wants to go beyond cost savings. It is building a system to link the AI spending that went into shipping a given feature with the value that feature delivered to customers. That system is not finished yet. The next question, alongside the drop in daily costs, is what customer value that spending delivers.