KT sets October 30 launch for AI platform that selects models and combines bills

KT is giving businesses one place to choose AI models, track spending and settle their bills. The company introduced Token Factory in Seoul on October 27, 2026, and scheduled its corporate launch for October 30. It will start with 73 models and include automatic selection based on the task, desired quality and cost.

The service combines model access behind a single application programming interface, or API—the software connection applications use to send requests. Instead of maintaining separate subscriptions and tracking each service independently, companies can manage model use, expenses and payments together. KT says customers will be able to check usage across models and pay in Korean won through unified billing.

A model catalog with two supply routes

The initial lineup divides into seven models provided through KT’s domestic infrastructure and 66 global models. KT named its Mideum K 3.0, Upstage’s Solar Pro 4 and Motif Technologies’ Motif 3 among the domestic-infrastructure offerings. The broader catalog includes OpenAI’s GPT series, Anthropic’s Claude and models from Mistral AI, DeepSeek, xAI and Alibaba Cloud.

KT has partnered with model-distribution platform OpenRouter to widen its supply. It plans to increase Token Factory’s catalog to more than 500 models by the end of 2026. That is a rollout target, not the opening-day selection. The partnership also includes plans to distribute models operated on domestic infrastructure to overseas customers through OpenRouter.

Choosing a budget, not just a model

Token Factory’s Auto Model Router makes the model choice on a customer’s behalf. KT says it analyzes the question’s type and difficulty, the quality required and the cost. Users can set their preferred balance between quality and spending rather than name a model themselves. Prioritizing quality steers selection toward a high-performance model; prioritizing savings favors a cheaper one.

The system also compresses prompts by removing unnecessary content before processing. Kim Joon-seok, head of KT’s AI Agentic Lab, put the potential cost reduction from compression at about 20%, with routing to task-appropriate, lower-cost models contributing roughly another 5%. Separately, a fallback feature switches a failed request to another model designated in advance. That provides a backup route rather than requiring the customer to reconnect manually.

Up to 25%KT’s claimed token-use reduction

KT says its routing and prompt-compression technology can reduce token usage by up to 25%. This is a company-supplied performance claim.

KT will run models as well as bill for them

KT’s business model goes beyond collecting access fees for outside services. At the briefing, executive Lee Jin-hyung said the company would generate revenue by running models on graphics processing units in its own data centers. That is inference: operating a model to process requests and produce answers. Lee said supplying that computing directly gives KT more flexibility over margins than brokerage alone.

For billing, KT is adapting the detailed usage records used in telecommunications. Its Token Detail Records log which models a customer uses and how many tokens they consume. Tokens are the units models use to process input and generate output. Those records underpin consolidated charges rather than leaving businesses to reconcile separate model-provider bills.

Two opening plans

• Standard charges for actual token use and supports postpaid settlement.
• Enterprise uses postpaid contracts, with options for dedicated networks, private environments and service-level agreements setting out service commitments.

KT says the platform will not store customer data or use it for training. It also describes safeguards that detect personal data, sensitive information and harmful expressions in inputs, and check generated answers. The initial service is for businesses; KT plans to add corporate management functions by December and extend access to individual developers and general users in March 2027.