A LegalOn Technologies mintegy 65 százalékkal csökkentette a napi fejlesztési költségeit azzal, hogy optimalizálta az OpenAI modelljeinek használatát. A vállalat korábban korlátlan hozzáférést biztosított a fejlesztőknek a GPT-5.5 modellhez, ám a növekvő kiadások miatt új modelkiválasztási irányelveket vezettek be.
A mérnökök a feladatok nehézsége alapján választanak a GPT-6 különböző verziói közül. A legegyszerűbb kódolási feladatokat a GPT-6 Luna végzi, a közepes tervezési és elemzési munkákat a GPT-6.1 Sol kapja, míg a legösszetettebb rendszertervezést a GPT-6 Astra modellre bízzák. Emellett részlegenkénti és egyéni költségkereteket is bevezettek.
Yuta Tokitake, a cég vezető mérnök-menedzsere szerint a cél most egy olyan új mutató kidolgozása, amely a fejlesztési sebesség helyett az ügyfélérték alapján méri az AI-befektetések megtérülését. A vállalat a jövőben a felvételi folyamat során is kiemelten figyeli majd a jelentkezők AI-készségeit.
Az eredeti szöveg (OpenAI)
LegalOn cut costs by 65% while maintaining development speed by matching Astra, Sol, and Luna to tasks and managing budgets strategically.
LegalOn Technologies offers Professional AI globally, using AI to support legal work and other key business functions. By combining domain expertise with AI, it enables people to focus on complex judgments and decisions. Its goal is AI-driven management that improves the quality and speed of business decision-making.
This approach extends beyond its products to making the organization itself AI-native. The company integrated Codex into its development process and expanded adoption across the organization through day-to-day use.
As adoption took hold, a new challenge emerged: cost control. Unlimited use of high-performance models could drive up spending, while blanket restrictions risked undermining the productivity gains AI had enabled.
How could it reduce costs without slowing development? LegalOn Technologies addressed this challenge by selecting among GPT‑6 (Astra, Luna) and GPT‑6.1 Sol based on task complexity and development stage, and aligning budgets with each business’s stage of growth. The company halved costs while maintaining development speed.
The company initially gave developers unlimited access to its main model, GPT‑5.5 in Fast mode. They expanded its use to design, implementation, and everyday work, learning through experimentation how best to divide responsibilities between people and AI.
Continuing to use high-performance models without limits, however, would inevitably exceed the annual budget. The company's AI-powered Development CoE (AID CoE) began developing guidelines for model selection. AID CoE tested and monitored models, while managers shared its findings with their teams. This enabled each engineer to independently choose the most suitable model for each task.
The new guidelines shifted the company from using the highest-capability model for every task to choosing the right model for the job. Teams started with a lightweight model and moved to more capable models as task complexity increased.
To support governance, the company also introduced system controls through administrator settings, with monthly usage limits for departments and individuals. AID CoE monitors usage and adjusts limits as business needs require.
Internal testing under these controls helped make the selection criteria more specific. GPT‑6 Luna, the lightest model, handles code implementation with clear requirements; GPT‑6.1 Sol supports standard design, data analysis, and document preparation; and GPT‑6 Astra, the most capable model, handles advanced judgment, including architecture design. Matching models to tasks gradually became a shared practice across teams.
Everyday execution and implementation; running everyday automations; delegating relatively simple tasks and analysis; code implementation, mainly as a subagent.
Standard design and analysis; relatively simple software design; routine numerical analysis and document creation; tasks requiring shorter completion times than with Luna.
Complex analysis, design, and orchestration; complex analysis and document creation; complex software design; coordinating agents as the orchestrator.
Alongside model selection, the company began restricting Fast mode by default, allowing individual requests only when needed. Although the change raised concerns about development speed, teams used approaches such as running tasks in parallel to maintain performance and transition smoothly.
It also introduced budget caps for AI spending at the department, group, and individual levels. The established LegalOn business was asked to improve cost efficiency by up to approximately 20%, while new businesses in the launch phase received generous budgets to encourage active use of AI.
Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies, explains that new businesses prioritized business speed over cost efficiency. The aim was to use AI exte
