MIDM K 3.0 Pro vs GPT-6.1 Sol: A Seventeenfold Price Gap and a Missing Score

On the afternoon of November 11, 2026, a million input tokens plus a million output tokens cost 70 cents on MIDM K 3.0 Pro and $12.00 on GPT-6.1 Sol. That is the entire case for KT's new model, stated as a number, and it is a strong case: the Korean telecom's 120-billion-parameter mixture-of-experts flagship, released on October 29, 2026 alongside the Token Factory platform that serves it, undercuts OpenAI's mid-tier workhorse by roughly seventeen to one on that mix, and by eighteen to one on output tokens alone. There is also a second number, and it is zero: the count of independent benchmark results anyone has published for MIDM K 3.0 Pro. Artificial Analysis carries a catalogue entry for KT's previous model, Mi:dm K 2.5 Pro, and left its intelligence score blank; for this one it has no entry at all. So this is not a comparison of two measured models. It is a comparison of a measured model against a price.

Everything below is either KT's own published material — the Token Factory model page, its pricing table, its service notices — or OpenAI's, or an Artificial Analysis measurement read on November 11, 2026. Where a figure is the vendor's and nobody has rerun it, the sentence carrying it says so.

What KT actually shipped, and on what day

The naming needs one sentence of unpacking, because searching for it is harder than it should be. KT's model family is written Mi:dm in Latin script and 믿음 in Korean; the API identifier is kt/midm-k-3.0-pro; this blog and most catalogue listings render it MIDM K 3.0 Pro. All three names are the same product. Korean trade press had the model a fortnight early — iNews24 reported on October 13 that the 120-billion-parameter "믿음 K 3.0 프로" would ship that month — and the Token Factory model page now carries a release date of 2026-10-29, the same day the platform's own service notice went up and one day before the launch discount started.

The architecture is more interesting than the parameter count suggests. It is a hybrid stack: Mamba layers, mixture-of-experts layers and ordinary attention, mixed rather than stacked in the usual transformer way. Of the 120 billion parameters, 12 billion are active per token, which is the ratio that makes a $0.20 input price arithmetically possible on a single provider's GPUs. KT also publishes the quantization scheme, which almost nobody does: the MoE experts run at NVFP4 while the Mamba projections and the shared expert stay at FP8. Context is 261,795 tokens with a maximum output of 235,615 — an unusual pairing, since most vendors cap output far below context. Input and output are text only; there is no image path.

• Total and active parameters — 120B total, 12B active per token

• Architecture — hybrid Mamba + MoE + attention; NVFP4 experts, FP8 Mamba projection and shared expert

• Context and output cap — 261,795 input, 235,615 output

• Modality — text in, text out; no vision, no audio

• List price — $0.20 per million input, $0.80 per million output, $0.05 cache read

• Price in force today — $0.14 / $0.56 / $0.04, a 30% launch discount running October 30 to November 30, 2026

• Release date — October 29, 2026, on KT Token Factory

• Independent score — none published

The gap, and the day it narrows

The discount is the first thing to get right, because a comparison written against $0.14 and read in December is wrong. KT's notice board lists the promotion as running 10/30 through 11/30, on seven of its own hosted models, tied to the Token Factory launch. On December 1 the same tokens cost $0.20 and $0.80 — still a tenfold and twelvefold gap against GPT-6.1 Sol, still the cheapest serious flagship price on the board, but not the number a reader will see quoted in this week's posts.

Against that, GPT-6.1 Sol's rate card has its own shape that a single blended number hides. The base rate is $2.00 input and $10.00 output, with cached input at $0.10 and cache writes at $2.50. Above 272,000 input tokens the whole request reprices — $4.00 input, $15.00 output — and because Sol's context window is 1,050,000 tokens, that tier is reachable in ordinary work rather than being a curiosity. MIDM K 3.0 Pro has no published long-context step and a context window a quarter the size, so the two models do not even compete for the same documents. If your prompts run past 261,795 tokens, this comparison ends before it starts.

What you get for the extra money on the OpenAI side is not a mystery. GPT-6.1 Sol carries a reasoning-effort ladder from low through max, a documented knowledge cutoff, structured outputs, and the ordinary ecosystem of SDKs and agent harnesses that a two-month-old model from a Korean telecom does not yet have. MIDM K 3.0 Pro exposes the reasoning parameters and structured outputs on its catalogue listing, so the API surface is not the problem; the tooling around it is simply younger.

The number nobody has published

Here is the honest shape of the evidence. Artificial Analysis runs a fixed evaluation suite and publishes an Intelligence Index on a numbered revision; as of this reading the revision is v5.0.1, and GPT-6.1 Sol scores 55 on it, at a measured cost of $1.27 per index task and 56 output tokens per second, with 153 seconds to the first answer token on reasoning-heavy items. Those are independent measurements, and they are the ones to trust over any vendor slide.

MIDM K 3.0 Pro has no row. Not a low score, not a provisional score — no row. KT has not published benchmark numbers for it either, at least not in any form reachable from outside Korea on the day this was written; the model page is a specification sheet, a price, and a live latency chart, and that is all. This matters more than it usually would, because the price argument is doing all the work. A model at one-seventeenth the price that is within a few points of GPT-6.1 Sol is an obvious buy. A model at one-seventeenth the price that is fifteen points behind is a different product for a different job, and there is currently no way to tell which one KT has shipped.

The one performance signal that does exist is KT's own, and it contradicts itself in a way worth seeing. The provider table on the model page reports 19.98 output tokens per second and 0.67 seconds end-to-end latency for the single KT endpoint. The three-day median chart directly above it reports an average of 62.93 tokens per second, 0.27 seconds to first token and 1.47 seconds end-to-end. Both are labelled as measurements of the same endpoint; they differ by a factor of three. That is not a scandal, it is a vendor dashboard mid-build, but it means the only throughput figure available for this model should be treated as "somewhere between 20 and 63 tokens per second" until somebody outside KT measures it.

Worth noting alongside that: the model page states plainly that single-provider models do not get automatic routing on KT's own platform. MIDM K 3.0 Pro is served by KT alone. Whatever redundancy you want around it, you build yourself.

Where the cheap model is the right call

Three workloads make the price gap decisive regardless of the missing score, and none of them needs a frontier model.

• Classification and extraction at volume — sorting tickets, tagging documents, pulling fields out of Korean-language forms. A model that is 85% as good at this costs a tenth as much, and the throughput chart's ambiguity does not matter because the work is batched.

• Korean-first product surfaces — KT built this model for Korean, and the Token Factory catalogue is explicitly positioned as domestic hosting on Korean infrastructure with KRW billing. If your users write Korean and your compliance team wants the tokens processed in Korea, that is a requirement no American lab currently meets at this price.

• Draft-and-verify pipelines — let MIDM K 3.0 Pro produce the first pass and let a stronger model check it. At $0.56 per million output tokens the first pass is nearly free, and the second pass only sees the fraction that needs review.

The mirror image is just as clear. Anything that needs one million tokens of context, image input, a measured capability floor before you commit, or a vendor who will answer a support ticket in your timezone this quarter belongs on GPT-6.1 Sol. The honest summary is that these two models are not really substitutes; they occupy different price bands and one of them has no published evidence at all.

Holding both without betting on either

The practical problem with an unmeasured cheap model is not that it might be bad — it is that you find out in production. That is the case automatic failover exists for: point the cheap path at MIDM K 3.0 Pro, keep a proven model behind it, and let a timeout or a 5xx resolve to the second model instead of to your user's error page. The same routing policy can send the long-context slice of your traffic somewhere else entirely, since MIDM K 3.0 Pro's 261,795-token window will reject it anyway.

Worth being precise about what OrcaRouter does and does not carry here. GPT-6.1 Sol is in our catalogue at OpenAI's published list price with no markup added, so the $2.00 and $10.00 above are what lands on the bill rather than a reseller's version of them, and the model is reachable on the same key and the same endpoint as the other 200-plus models we route, which is what makes an A/B against your own traffic a model-string change rather than a second contract. MIDM K 3.0 Pro is not in our catalogue today; it is served by KT's own platform, and that is where you would call it. We would rather say that than imply otherwise.

What that leaves is a decision you can actually make this month. If the workload is Korean, high-volume and forgiving, the price gap is large enough to justify a two-week trial on KT's own key, with the discount running until November 30 as a reason to start now rather than in December. If the workload is long-context, multimodal, or needs a number before procurement signs, GPT-6.1 Sol is the measured option and the seventeenfold premium buys evidence as much as capability. And if the answer changes in the next month it will change for one reason: somebody runs MIDM K 3.0 Pro through a public evaluation suite and publishes the result. Until that happens, the most useful thing anyone can say about this matchup is the thing this page has said — one of these models has a score, and it is not the cheap one.