Kimi K3 vs DeepSeek V4 Pro: Doing the 17x Price Arithmetic

Moonshot prices Kimi K3 at $3 per million input tokens and $15 per million output. DeepSeek prices V4 Pro at roughly $0.435 and $0.87. That is about seven times on input and seventeen times on output, between two models that both claim frontier-adjacent capability with 1M-token context windows.

A gap that large is not a rounding error you can ignore, and it is also not automatically a reason to pick the cheap one. This article is the arithmetic that decides it, because "which is better" is not answerable and "when does the premium pay back" is.

Start with cost per task, not cost per token

Per-token pricing is a supplier-side unit. What you actually consume is tasks, and a task costs turns multiplied by tokens per turn multiplied by price per token.

There is a useful published data point here. Artificial Analysis measured K3 on their private long-horizon knowledge-work evaluation at an overall Elo of 1547 — second only to Claude Fable 5 — with a cost per task of $0.94, against $1.04 for GPT-5.6 Sol. They also noted K3 used 21% fewer output tokens than Kimi K2.6 while scoring 732 Elo points higher.

That is the mechanism by which an expensive model can be competitive: it spends fewer tokens and fewer turns. It is not a claim that it beats a seventeen-times-cheaper model on cost, but it means the naive multiplication overstates the gap.

The break-even, made concrete

Suppose a task takes DeepSeek V4 Pro N turns and Kimi K3 M turns, with similar output length per turn. K3 is cheaper overall when M/N is below about 1/17 — which is to say, essentially never on turn count alone. Seventeen times is too large a gap to close through efficiency.

So the premium cannot be justified on token economics. It can only be justified on outcomes, and that changes the question to: what fraction of your tasks does the cheap model fail outright?

Work it through. If V4 Pro completes 90% of your tasks acceptably and K3 completes 97%, the premium is buying you seven percentage points. Whether that is worth seventeen times the token cost depends on what a failure costs — measured in engineer minutes spent reviewing, correcting or debugging a bad diff.

Put a number on that. An engineer at, say, £70 an hour spending twenty minutes untangling a wrong change costs about £23. If a task costs $0.05 on V4 Pro and $0.85 on K3, the premium is roughly $0.80 — around £0.60. The premium pays for itself if it avoids one failure per forty tasks. At a seven-point difference in success rate, it avoids roughly one per fourteen.

Plug in your own rates and your own measured pass rates. The structure is what matters: token cost is almost always small relative to review time, which is why "the cheap model is seventeen times cheaper" is rarely the decisive fact people assume it is.

Where the capability gap actually is

The gap is not uniform, and that is the part the arithmetic above cannot capture.

K3 sits around 57 on the Artificial Analysis Intelligence Index, the highest of any open-weight model. V4 Pro sits in the mid-forties. K3 ranked first on Arena's blind Frontend Code voting at 1,679 points, ahead of Claude Fable 5. Moonshot reports 91.2 on BrowseComp and 1668 Elo on GDPval-AA v2, and claims K3 beats Claude Opus 4.8 max and GPT-5.5 high on most tasks while trailing Claude Fable 5 and GPT-5.6 Sol.

But DeepSeek's own report puts V4-Pro-Max at 80.6% on SWE-bench Verified, 93.5 on LiveCodeBench, 95.2 pass@1 on AIME 2025 and a Codeforces rating of 3206. On self-contained problems with a verifiable answer, V4 Pro is not seventeen times worse. On many of them it is not worse at all.

Where the gap opens is ambiguous, exploratory, judgement-heavy work: reading unfamiliar code, deciding what a vague ticket means, holding a plan across a long session, producing UI that a human will look at and prefer. Reporting on V4 Pro consistently notes weakness in factual recall and in multi-step reasoning where no intermediate step can be checked.

Two more asymmetries

Caching. K3 charges $0.30 per million on cached input against $3.00 on a miss — a 90% discount. If you send a large stable prefix on every call, the input side of the gap narrows sharply. Design your prompts so the stable part comes first.

Licence. V4 Pro is MIT. K3 ships under a custom licence: attribution above roughly 100 million monthly active users or $20 million monthly revenue, and a separate agreement required for Model-as-a-Service businesses above $20 million in any twelve-month period. If you resell inference, this is not a cost question at all.

The rule that falls out

Do not choose. Route.

1. Send everything to the cheap model first. On most codebases it clears the large majority of tasks.
2. Define an escalation trigger — a failing test after two attempts, or a task tagged as exploratory.
3. Escalate that residue to K3.
4. Measure the escalation rate monthly. It is the single number that tells you what your model strategy costs.

Both models are served over OpenAI-compatible endpoints, so this is routing logic rather than a rewrite. Teams that do it typically find the expensive tier handles a small minority of requests and a large minority of the value — which is a far more actionable finding than a verdict on which model is better.