LegalOn Cut Codex Costs by About 65% by Matching Models to Tasks!
Hi everyone, it's Shii-chan!
OpenAI News
openai.com
Today I found a case study about managing AI coding costs, and I want to share it with you.
What was announced?
OpenAI News published a case study on LegalOn Technologies. They reduced estimated daily Codex costs by about 65% while keeping development speed.
LegalOn offers Professional AI globally for legal work and other business functions. It built Codex into its development process and spread it across the organization.
Why it matters
As adoption grew, a new challenge appeared: cost control.
- Unlimited use of high-performance models can drive spending up
- Blanket restrictions can undermine the productivity gains
Many teams hit this wall as AI coding spreads.
What changes
LegalOn first gave developers unlimited access to its main model, GPT-5.5 in Fast mode. Then it changed course with three measures:
- Guidelines for choosing models by task complexity
- Fast mode restricted by default (individual requests allowed when needed)
- Budget caps at the department, group, and individual levels
The result: estimated daily costs fell by about 65% compared with GPT-5.5!
Dive Deep
Matching models to tasks
The company's AI-powered Development CoE (AID CoE) tested and monitored models and built selection guidelines. Teams start with a lightweight model and move up as complexity grows.
- GPT-6 Luna: the lightest model. Code implementation with clear requirements, everyday automations, implementation as a subagent
- GPT-6.1 Sol: standard design and analysis, relatively simple software design, routine numerical analysis and document creation
- GPT-6 Astra: the most capable. Complex analysis and design, architecture design, coordinating agents as the orchestrator
Governance and budgets
Administrator settings set monthly usage limits for departments and individuals, and AID CoE adjusts them as business needs change.
Budgets follow each business's stage: established businesses were asked to improve cost efficiency by up to about 20%, while new businesses in the launch phase got generous budgets to encourage active AI use.
Restricting Fast mode raised concerns about speed, but teams kept up using approaches such as running tasks in parallel.
What's next
- Building its own ROI metric that links AI costs to customer value, with one feature release as the unit
- Planning a knowledge base to share practical know-how
- Starting to emphasize AI skills in hiring
Wrap-up
- Stopped using the most capable model for everything, and split work across Luna, Sol, and Astra
- Restricted Fast mode by default and kept speed with practices like parallel execution
- Tuned budgets per business stage, investing more in businesses it wants to grow
- Estimated daily costs dropped about 65% versus GPT-5.5
This one is for engineering managers worried about AI coding bills!
