DeepSeek V4.1 Flash vs Kimi K3: Reading the Claim, Then the Evidence

DeepSeek V4.1 Flash arrived with a specific, checkable claim: that it beats Kimi K3 on cyber and coding benchmarks — and on the rows the vendor selected, it does. Terminal-Bench 2.1: 90.6 against 88.3. CyberGym: 88.1 against 80.0. DeepSWE v1.1: 74.2 against 67.5. NL2Repo-Bench: 65.4 against 58.0. Every one of those DeepSeek figures comes from DeepSeek's own harness at DeepSeek's own reasoning effort, and every one of the Kimi K3 figures compared against it came from somewhere else. That is not a reason to dismiss the claim. It is the reason the claim needs unpacking before you buy anything on the strength of it.

Here is what the two models' evidence bases actually look like, what happens when you put them on one scale, and the one number that survives every version of this comparison.

Two evidence bases that barely overlap

This is the unusual part of this matchup, and it explains most of the disagreement you will find online.

DeepSeek published a launch benchmark table. Terminal-Bench 2.1, DeepSWE v1.1, CyberGym, ProgramBench, NL2Repo-Bench, Automation-Bench — all vendor-reported. They are the rows DeepSeek chose to run, on harnesses DeepSeek configured, with DeepSeek setting the reasoning effort. Worth knowing before you line them up against anything: Artificial Analysis found that running these suites at max reasoning effort consumes roughly 2.5x more output tokens than a mid-tier setting while recovering most of the accuracy below it. A vendor table at max effort is a ceiling measurement, not a typical one.

Kimi K3's numbers are mostly not Moonshot's. Artificial Analysis ran its own harness and published per-benchmark results on a scale comparable across every model it tracks: GPQA Diamond 93.5, Humanity's Last Exam 46.9, Long-Context Recall 88.7, SciCode 59.5, Terminal-Bench 2.1 85.0 — plus a composite index. Those are a third party's measurements.

So "DeepSeek beats Kimi K3 on coding" resolves to something narrower than it sounds. On DeepSeek's harness, DeepSeek's model wins the coding rows DeepSeek selected. That is a real signal — vendors do not usually publish tables they lose — but it is a different class of evidence from an independently run suite, and the two should never appear in the same column without a label.

On one scale, the gap is four points

Put both on Artificial Analysis's Intelligence Index and the picture is tighter than either side's marketing suggests. Kimi K3 scores 44. DeepSeek V4.1 Flash scores 40.

Four points. On the per-benchmark rows, Kimi K3's lead is consistent and concentrated where long-horizon reasoning shows up: Humanity's Last Exam 46.9 against 36.8, SciCode 59.5 against 51.9, GPQA Diamond 93.5 against DeepSeek's vendor-reported 90.9, Long-Context Recall 88.7 against 84.0. Kimi K3 also ranks second on Artificial Analysis's board of 113 models, against DeepSeek V4.1 Flash at sixth.

Those are genuine, independent, reproducible gaps. No price advantage closes them for a task that depends on them.

They are also four points — the same four-point gap that separates DeepSeek V4.1 Flash from Grok 4.6, and smaller than the five-point gap separating it from Qwen3.8-Max. A four-point spread on a composite index is a narrow band. It tells you Kimi K3 is the stronger model and it does not begin to tell you that the strength shows up in your workload.

The number that survives every version of this comparison

Cost per Intelligence Index task, measured by the same organisation that produced the index: $0.27 for DeepSeek V4.1 Flash, $2.00 for Kimi K3.

7.4x. That is the largest, most robust, most cross-checkable difference between these two models, and unlike the benchmark rows it does not depend on whose harness ran it — the token counts are the token counts.

It also scales with how agentic your workload is. The $2.00 figure is from a mixed suite; on Vals AI's agentic coding bench the reported gap widens to roughly 40x, $0.41 per task against $17.59, with the DeepSeek run finishing in about 15 minutes against roughly 90. The further you get from a single prompt and response toward a long tool-calling loop, the more the two pricing models diverge — because that is where caching and output volume dominate the bill.

• Rate — DeepSeek V4.1 Flash $0.15/$0.60 per 1M off-peak ($0.30/$1.20 peak, weekday 01:00–04:00 and 06:00–10:00 UTC) vs Kimi K3 $3.00/$15.00 per 1M, flat

• Cache discount — 98% vs 90%; cache read $0.003 per 1M vs $0.30 per 1M

• Context — 1,048,576 tokens both, and neither applies a long-context surcharge

• Throughput — 214.4 tokens/sec vs 34.7, a 6.2x gap

• Verbosity — 250M output tokens across the index vs 160M

• Parameters — 552B total, 8B active on prefill and 16B on decode, vs 2.8T total with 16 of 896 experts active per token

• Sampling — full temperature / top_p / seed surface vs none; Kimi K3 exposes reasoning_effort only and thinking cannot be disabled

• Licence — MIT vs Modified MIT, which adds branding terms above 100M monthly users or $20M monthly revenue

Where the cheap side is genuinely worse

Two costs sit on the DeepSeek side of the ledger, and both are easy to miss.

Verbosity. DeepSeek V4.1 Flash emitted 250M output tokens across the Intelligence Index against Kimi K3's 160M. Output bills at 4x input here. A model that is twenty times cheaper per token but talks half again as much is not twenty times cheaper — which is exactly why the honest figure settles at 7.4x and not the 20x the rate card implies. DeepSeek is the most verbose model in this comparison; Grok 4.6 is the least, at 94M.

Speed. The rate card says DeepSeek is cheaper. The clock says it is 6.2x faster too — 214.4 tokens/sec against Kimi K3's 34.7. Kimi K3's thinking mode cannot be turned off and its reasoning effort defaults to max, so there is no cheap, fast path to fall back to for the easy calls. You buy frontier reasoning on every request whether the request needs it or not. On a hard coding task that is the product. On a classification pass it is waste.

What the traffic says, which no benchmark captures

Our own routing data adds the one signal neither vendor nor analyst can produce. Over the last seven days, Kimi K3 moved about 3.04 billion tokens through OrcaRouter. DeepSeek V4.1 Flash moved about 59.1 billion — roughly nineteen times as much, in a model that has been live for under a week.

Part of that is structural. DeepSeek is routing all V4 Pro requests to V4.1 Flash and billing them at Flash rates until V4.1 Pro ships, so a large installed base was migrated onto it without anyone choosing anything. But the resulting volume is still the clearest available evidence that an enormous number of production workloads were already being served well enough by the cheaper model — and that is a verdict benchmarks cannot give you, because benchmarks measure what a model can do and traffic measures what teams were willing to put in front of paying users.

Choosing between them

Both sit behind the same OpenAI-compatible endpoint on OrcaRouter, which is what makes the split practical rather than a migration project: route the long-horizon coding session where Kimi K3's reasoning lead is worth paying for, and the high-volume agentic and knowledge-work traffic where the two are within four index points, to DeepSeek V4.1 Flash. Because the gateway passes the provider rate through with zero markup, the per-token cost you see is the vendor's own — worth checking twice a day on the DeepSeek side, since its peak windows double the rate on weekday mornings UTC.

The verdict follows the evidence rather than the launch table. If your work is one hard reasoning call — a scientific question, a novel algorithm, a document you cannot afford to misread — Kimi K3 is the stronger model, the independent gaps are consistent, and the price is secondary. If your work is a coding agent, a long tool-calling loop, or anything with a large cached prefix replayed at volume, the two are close enough on capability that cost becomes the deciding criterion, and there the gap runs from 7.4x to roughly 40x depending on how agentic the task is.

DeepSeek's own benchmark table overstates how clean that win is. The price does not.