LegalOn: How "Tiered Model Selection" Slashed AI Development Costs by 65%

Overview

LegalOn Technologies is a tech startup providing professional AI services globally, dedicated to combining legal domain expertise with artificial intelligence to help clients focus on complex judgment and decision-making. As the company deepens its development into an AI-native organization, seamlessly integrating Codex into development workflows while controlling costs has become a core challenge it faces.

"Applying too many restrictions through rules and budgets can undermine the organization's momentum. What we need is a flexible operating model that can effectively control risks while ensuring the team's required pace of progress." — Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies

Tiered Model Selection: The Key Strategy for Balancing Cost and Performance

Facing budget pressures from high-performance models, LegalOn's AI-Driven Development Center of Excellence (AID CoE) did not adopt a one-size-fits-all restriction approach. Instead, it formulated a refined tiered model selection guideline. The team flexibly switches between GPT-6 series models (Astra, Luna) and the GPT-6.1 Sol model based on task complexity and development phases.

The core of this strategy lies in matching on demand: the lightest models handle clear requirement code implementation, mid-level models process routine design and documentation, and the most capable models focus on advanced tasks such as architecture design. Meanwhile, the company defaults to limiting fast mode, allowing its use only upon individual request when truly necessary.

Specific measures include:

• GPT-6 Luna: Handles daily execution, simple task analysis, and serves as a sub-agent for code implementation.
• GPT-6.1 Sol: Used for standard design, data analysis, and document writing requiring shorter completion times.
• GPT-6 Astra: Takes on complex analysis, software design, and coordination tasks across multiple agents.

By introducing system controls to set monthly usage caps for departments and individuals, and allocating differentiated budgets according to different business phases, LegalOn has achieved significant cost optimization while maintaining its original development speed.

From Cost Control to Value Measurement: Return on Investment Beyond Efficiency

After initially achieving cost reductions, LegalOn did not stop there. Senior Engineering Manager Yuta Tokitake pointed out that simply accelerating development speed does not equate to the success of AI investment; the key lies in whether development speedups have truly translated into customer value.

To this end, the company is building a new measurement indicator system, shifting evaluation criteria from development speed and usage volume toward customer value. The team plans to treat each feature release as an independent unit, linking the customer value it generates with the actual AI costs invested, thereby clearly presenting the return on investment.

Additionally, LegalOn is committed to transforming individual engineers' practical experience into organizational capability. By establishing a knowledge base to consolidate best model combinations and design patterns, the company is promoting the importance of AI skills in hiring and continuously optimizing the balance between governance and team autonomy. These efforts aim to build a lasting competitive advantage for the AI era.