Key Takeaways
- Verified about 90% reduction in AI coding operating costs
- The goal: total cost per task, not just token unit price
- Are all requests being routed to a single high-performance model?
- This week: tally your internal AI requests by task difficulty
At a press briefing on November 12, 2026, LG Uplus disclosed internal verification results showing that using its proprietary AI model for coding tasks can cut operating costs by about 90% compared to commonly used models. At the same event, the company announced plans to launch an enterprise AI cost-reduction business starting in 2027, through partnerships with other companies and its own infrastructure buildout. The goal isn't simply to lower the per-token price, but to reduce the total cost of having AI complete a given task.
How did LG Uplus cut AI costs?
The technology LG Uplus presented at its 'AI Token Optimization Study' session spans four areas.
- Model RoutingA technique for selecting the most suitable AI model for a given task from among several options—using high-performance models for complex tasks and lighter models for simple ones. In LG Uplus's internal tests, this cut the cost of handling general queries by more than 25% and coding agent costs by more than 30%.
- Improved Computation ProcessingProcessing efficiency on the same infrastructure was improved by up to 4x.
- Model LightweightingThe company shrank its proprietary on-device AI model from 2.4–3GB down to about 800MB, while retaining over 95% of the performance of comparable cloud models. Using a technique where a smaller model generates responses and a larger model verifies them, token generation speed more than doubled, and two related papers were accepted at NeurIPS.
- AI Chip DiversificationLG Uplus is validating the use of FuriosaAI's NPUs alongside GPUs from Nvidia and AMD. The aim is to lower infrastructure costs by distributing computation across different chips depending on workload characteristics.
These optimization technologies are being applied to the AI call agent 'ixi-O,' the AI Contact Center (AICC), and internal task automation. Ixi-O initially ran LG AI Research's on-device model alongside Google Gemini in parallel, and later adjusted the usage ratio between its proprietary model and EXAONE, among others, to improve cost efficiency. LG Uplus is also handling AI model operations for 'AI for Everyone,' a collaborative service with Kakao slated for commercial launch by year-end.
"Dominance in the AI market has shifted from who can build the biggest model to who can deliver the business outcomes customers want most precisely and economically." — Sangyup Lee, CTO of LG Uplus
Implications for Korean Enterprise AX
CTO Sangyup Lee explained that as AI agents proliferate, even cheap per-model pricing can balloon total costs if unnecessary searches and analyses are repeated. When evaluating AI adoption costs, the right order is to first examine how many model calls are needed to complete a single task and which models are used—before comparing token price sheets. The 25% and 30% routing gains are internal LG Uplus test figures, so organizations should measure their own results the same way before applying them.
This week, try separating your organization's internal AI requests into complex and simple tasks and count how many fall into each category. If even the simple tasks are all being routed through a single high-performance model, that's the first place to consider applying model routing.
Source: LG Uplus "Cuts AI Coding Costs by 90%"... Pursuing Token Business Next Year
