MIDM K 3.0 Pro vs Grok 4.7: What $7.77 a Task Hides About a $0.56 Model

Artificial Analysis charges Grok 4.7 for one Intelligence Index task at a rate of $7.77. The model lists at $2.00 per million input tokens and $6.00 per million output, which sounds reasonable until you notice it burns 148 million output tokens getting through the suite, roughly twice the median, and takes 27 minutes per task to do it. KT's MIDM K 3.0 Pro, released October 29, 2026, lists at $0.20 and $0.80 with a 30 per cent launch discount on top, and nobody has run it through a public evaluation suite at all. Putting Grok 4.7 next to MIDM K 3.0 Pro is therefore not a contest between a good model and a bad one. It is a contest between a measured model that turns out to be expensive in a way its rate card does not advertise, and an unmeasured model whose entire argument is that it is cheap enough to find out about.

Two rate cards, one of them audited

The audited one first, because it is the more interesting of the two. Grok 4.7 shipped on September 21, 2026 as SpaceXAI's flagship for long-horizon coding and agentic work, with a 500,000-token context window, image input, and a reasoning-effort selector across four levels. On the highest of those settings, Artificial Analysis scores it 35 on the Intelligence Index — below the median of 40 for models in its price band. That is a genuinely awkward number for a flagship, and it is not the only one: at 64 output tokens per second it is slow, at 27.2 minutes per index task it is slower still, and the 148 million tokens it emits across the suite is more than twice the 72.8 million median. Verbosity is not a moral failing, but it is a line item, and that is how $2.00 and $6.00 becomes $7.77.

The unaudited one is a different shape entirely, and the shape is the point. KT built MIDM K 3.0 Pro as a hybrid stack — Mamba layers mixed with mixture-of-experts and ordinary attention — carrying 120 billion parameters of which 12 billion are active per token. Sparse activation is what makes $0.20 input arithmetically possible; you are not paying for parameters you did not use. The MoE experts are quantized to NVFP4 while the Mamba projections and the shared expert stay at FP8, which is a level of disclosure almost no vendor bothers with. Context is 261,795 tokens. Input and output are text only — no images, which is a real difference from Grok 4.7 rather than a footnote.

• List price — MIDM K 3.0 Pro $0.20 in / $0.80 out per million, against Grok 4.7 $2.00 / $6.00

• Price in force today — MIDM K 3.0 Pro $0.14 / $0.56 until November 30, 2026; Grok 4.7 unchanged

• Cached input — MIDM K 3.0 Pro $0.04 per million, against Grok 4.7 $0.50

• Context — 261,795 tokens against 500,000, with Grok 4.7 allowing 450,000 tokens of output

• Modality — text only, against text and image in

• Architecture — 120B total and 12B active, hybrid Mamba/MoE/attention, against an undisclosed parameter count

• Independent score — none published, against 35 on Artificial Analysis v5.0.1

• Measured cost per index task — not measured, against $7.77

The arithmetic nobody puts on a pricing page

Run a concrete job through both cards and a tenfold sticker gap turns into something larger and stranger. Take a call that reads 30,000 tokens of input and writes 1,500 tokens of output, repeated ten thousand times — a mid-sized extraction or review pipeline, not an extreme case.

On Grok 4.7 at list rates that is 300 million input tokens at $2.00, so $600, plus 15 million output tokens at $6.00, so $90. Six hundred and ninety dollars. On MIDM K 3.0 Pro at the discounted rate it is 300 million input at $0.14, so $42, plus 15 million output at $0.56, so $8.40. Fifty dollars and forty cents — roughly fourteen times less than Grok 4.7 for the same shaped job. At list rates from December 1 the same job costs $60 and $12, so $72, which is still about ten times less.

Now the part that makes the comparison honest rather than promotional: the Grok 4.7 figure assumes the model answers in the number of tokens you asked for. It does not, reliably. The 148-million-token reading against a 72.8-million median says that on hard work this model spends roughly twice what its peers spend to reach the same answer, and a rate card multiplied by an unknown token count is not a quote, it is an estimate. The $7.77 per index task is the closest thing to a real unit price either vendor has published, and there is no equivalent figure for MIDM K 3.0 Pro, because no one has run the suite. Estimate it from the rate card and a median token count and you get well under a dollar; estimate it from a model that overspends the way Grok 4.7 does and you get several. That spread is the whole problem with pricing an unmeasured model from a rate card.

Where the missing score actually bites

It would be easy to write this as "the cheap model wins", and wrong. Three things follow from the evidence as it stands.

The first is that Grok 4.7's 35 is not a ceiling you should assume MIDM K 3.0 Pro clears. A 120-billion-parameter model with 12 billion active is a small model by 2026 flagship standards, and small models are cheap for reasons. The plausible range for MIDM K 3.0 Pro runs from "close enough to 35 that the price decides it" to "twenty points behind and fine only for tagging" — and there is currently no way to narrow it from outside KT.

The second is that the price gap is wide enough to absorb a real capability gap and still win on economics. If MIDM K 3.0 Pro is fifteen points behind Grok 4.7, it is still the correct choice for a classification job where the failure mode is a mislabelled ticket. If it is fifteen points behind on multi-hour agentic coding, it is not a substitute at any price, because the work does not complete.

The third is that Grok 4.7's own cost profile makes it a worse default than its reputation suggests. A flagship that costs $7.77 a task and 27 minutes of wall clock is a specialist instrument. The teams getting value out of it are using it on the hard tail, not on volume, and the volume is where a model like MIDM K 3.0 Pro is aimed.

The test that costs fifty dollars instead of seven hundred

There is a way to answer the capability question without waiting for a public leaderboard, and the price gap is what makes it practical. Assemble two hundred real requests from your own traffic — the ones with known-good answers, not invented examples — and run them through both models. On MIDM K 3.0 Pro at the promotional rate that costs pennies. On Grok 4.7 at $7.77-equivalent token behaviour it costs a few dollars. What you get back is not an Intelligence Index score, but it is a better number than an Intelligence Index score, because it is scored on your distribution rather than on a synthetic suite.

Two caveats before anyone runs it. First, the discount ends November 30, 2026, so a test run in early December measures a model that costs 43 per cent more on input and output than the one you read about this week. Second, MIDM K 3.0 Pro is served by exactly one provider — KT itself — and KT's own model page states that single-provider models do not receive the platform's automatic routing. There is no second endpoint to fall back to if the first one degrades.

That last point is the one place where OrcaRouter is directly relevant to this matchup rather than incidentally. Grok 4.7 is in our catalogue at SpaceXAI's published list rate with nothing added on top, on the same key and the same endpoint as the other 200-plus models we route, so the $2.00 and $6.00 on the card are the numbers that reach the invoice and a vendor price change appears the same day rather than at the next billing cycle. MIDM K 3.0 Pro is not something we serve; it lives on KT's Token Factory platform, and that is the honest answer to "where do I get it". What a router does for a pair like this is let you keep the measured model as a fallback behind the unmeasured one — a request that fails, times out, or exceeds 261,795 tokens resolves somewhere that works instead of to an error, which is the difference between trialling a cheap model and betting a production path on it.

What would change this page

One publication would. If Artificial Analysis or any comparable evaluator puts MIDM K 3.0 Pro on a numbered suite and it lands anywhere near 30, the price argument becomes decisive for a large class of work and the Grok 4.7 comparison stops being interesting. If it lands in the low twenties, the model is a Korean-language workhorse and should be sold as one, which is roughly how KT is positioning it anyway.

Until then the defensible reading is narrower than either vendor would like. Grok 4.7 is a measured model with a real capability level — 35, just under the median — and a cost per task that makes its sticker price misleading by a factor of four or five. MIDM K 3.0 Pro is an unmeasured model with a sticker price that makes almost any capability gap survivable on volume work, and no evidence whatsoever about whether it clears that bar. The only cost-per-task figure either side has actually produced is Grok 4.7's $7.77, and it was produced by a third party rather than by the vendor.