KT's Token Factory Is Live: 73 AI Models, One Bill, and a 25% Token-Saving Claim

Korea's largest telecom company just entered the AI business from the billing side. KT unveiled its Token Factory AI platform at a press briefing in Seoul on October 27, 2026, and the service goes live for corporate customers today, October 30 — 73 AI models behind a single API, one unified bill in Korean won, and a proprietary model router KT says can cut token usage by up to 25%.

It's a bet that the next war in AI isn't about who builds the better model, but who makes models cheapest to actually use. KT's Park Sang-won, senior executive vice president and head of the AX Business Division, put it exactly that way at the briefing: "The AI market is rapidly shifting from a competition over who can build better models to one over who can support the easier and more efficient utilization of AI." Token Factory is KT's bid to own that layer in Korea — and, eventually, beyond it.

What launched today

Token Factory is a single-platform gateway to multiple AI models. Companies sign up once, call one API, and get usage monitoring, cost management, and payment handled in one place — no separate accounts or individual contracts per model provider.

The day-one catalog:

• 73 models at launch — 7 served through KT's own domestic infrastructure and 66 global models. Among the domestic offerings, KT named its own Mideum K 3.0, Upstage's Solar Pro 4, and Motif Technologies' Motif 3. The global side includes OpenAI's GPT series, Anthropic's Claude, and models from Mistral AI, DeepSeek, xAI, and Alibaba Cloud.
• 500+ models by November — via an official partnership with OpenRouter, the global AI model distribution platform. MoneyToday reports this is Korea's first such agreement with OpenRouter, and KT says it will also push models running on its domestic infrastructure out to overseas customers through the same partnership.
• Single API and unified billing in won — one integration point, consolidated token bills paid in Korean won, and telecom-grade usage records (see below).

The 500 figure is a November rollout target, not today's selection. That distinction matters: what goes live this morning is 73 models, all for corporate clients. KT says corporate management functions arrive by December, and individual developers and general users get access in March 2027 — which is also when subscription pricing arrives, according to the Korea Times.

Auto Model Router: set a budget, not a model

The platform's core feature is KT's in-house Auto Model Router, which picks the model for each request automatically. It analyzes the query's type, difficulty level, required quality, and cost — plus the range of models the customer has authorized — and routes simple work like summaries and translation to lightweight, low-cost models while reserving high-performance models for complex analysis.

The interface shift is the interesting part: instead of naming a model, customers set their preferred balance between quality and spending. Prioritizing quality steers selection toward high-performance models; prioritizing savings favors cheaper ones. There's also a fallback: if a request fails, the system switches to another model the customer designated in advance, rather than forcing a manual reconnect.

KT claims the router ranked first overall on RouterArena, a global benchmark that evaluates model-selection performance. That is KT's claim — reported by FNNews and others as the company's words — and I could not independently verify the leaderboard at write time, so treat it as vendor-supplied until third-party results surface.

The 25% claim, itemized

Every outlet at the briefing led with the same headline number: up to 25% token-use reduction. Here's how KT breaks it down, per Kim Joon-seok, head of KT's AI Agentic Lab:

• ~20% from prompt compression. KT's compression technology is tailored to Korean language characteristics — it strips repetitive and unnecessary expressions while preserving the question's meaning, reducing the number of input tokens.
• ~5% from model routing. Sending each task to a task-appropriate, lower-cost model shaves the rest off the total token bill.

Like the RouterArena claim, the 25% figure is company-supplied performance data, not an independently audited result. The composition is useful anyway: it tells you the savings come mostly from input-side compression (which benefits every model you call) and partly from routing (which benefits you only to the extent the router picks correctly). The compression is tuned for Korean — companies running mostly English workloads shouldn't assume identical numbers, a caveat KT itself flags by describing the technology as Korean-tailored.

The billing model: telecom DNA

This is where KT's 40 years of telecom billing experience shows. Token Factory adapts the detailed usage records used in telecommunications into Token Detail Records — per-customer logs of which models were used and how many tokens they consumed. Those records underpin consolidated charges instead of leaving businesses to reconcile separate provider bills — the same billing-plumbing direction as Mastercard's Agent Pay for Machines, where the payment rails, not the model, are the product.

Two plans at launch:

• Standard — pay for actual token use, with postpaid settlement. You pay for the tokens you consumed, full stop. (For how per-token pricing works in practice, see our DeepSeek V4.1 Flash pricing breakdown.)
• Enterprise — postpaid contracts with dedicated networks, private environments, and service-level agreements spelling out service commitments.

And here's the fine print you should notice: KT has not published a per-model token rate card. Coverage consistently says customers pay for "actual token use" and that billing is in won, but no briefing report includes per-model input/output rates for the 66 global models. If you're evaluating this against buying GPT or Claude access directly, the missing rate card is the first question to ask — the router's 25% savings are meaningful only relative to prices you can compare.

KT's real bet: selling tokens, not brokerage

The most revealing detail from the briefing wasn't about models at all — it was about money. Lee Jin-hyung, head of KT's AX division, told reporters that brokerage fees are not the main business model. Instead, KT will run models on GPUs in its own data centers and generate revenue by producing and selling tokens directly. In his words (via Chosun Biz), supplying computing directly gives KT "more flexibility over margins than brokerage alone."

That reframes Token Factory: it's not just a reseller UI on top of OpenRouter — it's an inference business. KT runs its in-house models and open-weight models from Korean data centers, which is also the data-sovereignty pitch: corporate data doesn't leave the country for those workloads, and KT's own guardrail technology screens inputs for personal data, sensitive information, and harmful expressions while checking generated answers.

A caveat worth stating plainly: the domestic-infrastructure claim covers KT's in-house and open-weight models. Coverage does not say the 66 global models — the GPTs, Claudes, and xAI models — are served from Korean data centers too; those presumably route to the providers' own infrastructure, with whatever data-handling terms those providers set.

The OpenRouter deal cuts both ways

The partnership with OpenRouter is doing double duty. Inbound: it takes KT's catalog from 73 to 500+ models in November, solving the coverage problem no single telco could solve alone. Outbound: KT plans to distribute models running on its domestic infrastructure — Mideum K 3.0 and friends — to overseas customers through OpenRouter, turning a domestic inference fleet into an export product.

KT is building an ecosystem around the platform too: collaborators named at the briefing include Eucal, CAST AI, and Timely, and the company says it's assembling participants through its "AI for All" consortium. The stated ambition is to grow Token Factory from a corporate AI platform into a global AI model distribution business — the telco as token wholesaler.

Token Factory vs. OpenRouter: how the gateways compare

The honest read: if you want manual control and developer access today, OpenRouter already exists. Token Factory's pitch is automation (routing + compression), Korean data residency, and won-denominated enterprise billing — aimed squarely at Korean corporates, not indie hackers.

What KT hasn't said yet

• No per-model rate card. "Pay for actual token use" is the billing promise, but per-model input/output prices were not disclosed at the briefing.
• 25% and RouterArena #1 are company claims. Every outlet attributes them to KT; neither has independent confirmation in the coverage.
• Korean-tuned compression. The headline savings number assumes workloads that look like KT's test set; English-heavy or code-heavy workloads may see different results.
• Data residency applies to part of the catalog. In-house and open-weight models stay in Korean data centers; the 66 global models route wherever their providers serve them.
• No consumer or developer access until March 2027. Today's launch is corporate-only; subscription pricing is still to come.

FAQ

Is Token Factory available to individuals?

Not yet. The October 30, 2026 launch is for corporate customers only. KT plans to extend access to individual developers and general users in March 2027, with subscription pricing arriving then.

Do I need separate accounts or API keys for each model?

No — that's the point of the platform. One signup, one API, and KT handles model selection, usage tracking, cost management, and payment across the catalog.

What does the "up to 25%" token saving actually cover?

Per KT's AI Agentic Lab: roughly 20% from Korean-tailored prompt compression and roughly 5% from routing tasks to lower-cost models. It's KT's claimed figure, not independently verified.

Does KT store my data or train on it?

KT says the platform will not store customer data or use it for training. It also describes guardrail technology that detects personal data, sensitive information, and harmful expressions in inputs and checks generated answers.

Can I use Token Factory outside Korea?

Nothing in the launch coverage indicates international availability at launch — it's a Korean corporate launch with won-denominated billing. KT's stated long-term goal is a global AI model distribution business, including selling domestic-model inference abroad through OpenRouter.

Key takeaways

• KT's Token Factory launched today (Oct 30, 2026) for Korean corporate customers: 73 AI models behind one API, one bill in Korean won.
• The in-house Auto Model Router replaces model-picking with a quality-versus-cost setting, plus failover to a pre-designated backup model.
• KT claims up to 25% token savings — ~20% from Korean-tuned prompt compression, ~5% from routing — but this is company-supplied, not independently verified.
• The real business model is inference, not brokerage: KT runs models on its own data-center GPUs and sells tokens directly, with domestic data residency for in-house and open-weight models.
• Watch the November OpenRouter integration (500+ models) and the March 2027 consumer launch — today's corporate rollout is only chapter one.

Published October 30, 2026. Sources: MoneyToday, The Korea Times, SuperPower Daily, AsiaE, SBS News, FNNews, MK, Chosun Biz, TheLec — all reporting from KT's October 27 press briefing in Seoul. Vendor claims are attributed to KT throughout.