Qwen 3.7 Max vs Kimi K3 vs DeepSeek V4: 34x Cost Gap (2026)

Three Chinese flagship families, one benchmark suite, and a 34x spread in what it costs to finish it. The capability gap between the best and worst of them is 13 index points. The bill gap is $2,365.

• Most capable: Kimi K3 (max), 57 on Artificial Analysis Intelligence Index v4.1
• Best value by a distance: DeepSeek V4 Flash 0731, 50 for $72.02 per index run
• Fastest: Qwen 3.7 Max, 206.4 output tokens/sec
• The surprise: the cheapest model in the group outscores both flagships it undercuts
• All four are on ofox today, so switching is a model-ID change

TL;DR: Which One Should You Pick?

Two ways to stop reading here. If your monthly token spend is under about $200, pick DeepSeek V4 Flash and move on, because none of the optimisation below is worth the afternoon. If you are running production agent loops above $2,000 a month, go straight to the eval-cost section, because that is where the 34x lives.

Quick Specs Comparison

All figures below are from the ofox model catalog, pulled live on 2026-08-03.

Three things in that table are worth pausing on.

The cache-read column is not a rounding difference. DeepSeek charges $0.0028 per 1M cached input tokens. Qwen 3.7 Max charges $0.50 for the same thing, which is 179x more, and K3 charges $0.30, which is 107x more. For a workload with a large stable system prompt read many times, the cached-read rate is the number that decides your bill, not the headline input rate.

Qwen 3.7 Max’s 64K output ceiling is the tightest of the four by a factor of six against DeepSeek and sixteen against K3. If your task is “read a large repo and emit a large artefact,” that ceiling bites before anything else does. Qwen 3.8 Max lifts it to 131K, covered in the Qwen 3.8 Max launch breakdown.

Only K3 has open weights and a frontier score. DeepSeek’s weights are MIT-licensed and downloadable, but its top scorer here is the cheap model, not the flagship. Qwen 3.7 Max has no open weights at all, though Alibaba has announced a Max-class open-weight release for the week of 2026-08-10.

How Do They Score on the Same Benchmark?

Artificial Analysis Intelligence Index v4.1, snapshot 2026-08-03. This is the only source in this article that measures all four models the same way, which is why the whole comparison is anchored on it.

The index is a composite of nine evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR.

Read the ordering, not the absolute numbers. This is a rolling leaderboard that gets re-scored as the index version and the model builds change, so “K3 first, Flash second” will age better than “57 and 50”. The four-point gap between Flash and Qwen 3.7 Max is inside the range a re-run can move.

What does not move is the direction of the surprise: the cheapest model in the group sits above two flagships, one of which is from the same vendor.

Three things would change this ranking, and all three are plausible inside a quarter:

• Qwen 3.8 Max gets scored. It is cheaper than 3.7 Max and Alibaba’s own table puts it well ahead of its predecessor. If it lands near K3 at $2.00 / $6.00, the value column shifts.
• A new DeepSeek build lands. The 0731 retrain moved Flash 10 points with no architecture change. Nothing says that was the last one.
• The index version changes. v4.1 added and reweighted evaluations; a v4.2 could reorder models separated by four points without any model changing at all.

That is the case for re-running your own evals quarterly rather than inheriting a ranking from an article, including this one.

Each vendor also publishes its own table, and those are not comparable with each other. Qwen’s launch table for 3.8 Max benchmarks against Claude and GPT with harnesses Qwen selected. Moonshot’s K3 launch table reports Terminal-Bench 2.1 at 88.3 against a leader of 88.8, which is a second place rather than a win. DeepSeek’s own Terminal-Bench figure for V4 Flash runs above the one Artificial Analysis measured for the same model. Those are different measurement setups, and mixing them into one table is how a comparison ends up meaningless.

What Does One Eval Run Actually Cost?

This is the number that reframes the whole comparison. Artificial Analysis publishes the total cost of running its own index on each model, so it is a real invoice for an identical workload rather than a sticker-price multiplication.

Derived from those four rows:

• K3 costs 33.8x what V4 Flash costs, for +7 index points.
• Qwen 3.7 Max costs 22.3x what V4 Flash costs, and loses by 4 points.
• V4 Pro costs 2.4x what V4 Flash costs, and loses by 6 points.
• K3 costs 1.5x what Qwen 3.7 Max costs, for +11 points.

One note on the price columns: Artificial Analysis lists DeepSeek V4 Pro at $0.435 / $0.87 while the ofox catalog rounds it to $0.45 / $0.88. The eval-cost figures come from Artificial Analysis and use its own numbers; the specs table above uses the catalog’s.

Only one of those four comparisons describes a trade. The other three describe a model that is worse and more expensive at the same time.

The K3 premium is at least defensible: you are paying 1.5x over Qwen 3.7 Max for a genuine 11-point lift, and if your work is bottlenecked on the hardest reasoning, that is a rational purchase. Paying 22x over V4 Flash to score four points lower is not a trade, it is an artefact of nobody re-running the numbers after 2026-07-31.

Why Is the Cheapest Model Beating Two Flagships?

Because the 0731 build was a retrain, not a new model, and it moved the score 10 points. DeepSeek V4 Flash 0731 kept the same 284B-total, 13B-active MoE architecture and came back scoring 50 against the previous build’s 40.

The second half of the answer is verbosity, and it cuts against DeepSeek.

• V4 Flash generated 210M output tokens on the index, the most of the four.
• Qwen 3.7 Max generated 100M, the least.
• So Flash emits 2.1x the tokens Qwen does for the same suite.

That is why the realised gap is smaller than the sticker gap. Qwen 3.7 Max’s output rate is 26.8x Flash’s ($7.50 against $0.28), but the actual index bills came out 22.3x apart. Flash gives back roughly a fifth of its price advantage by talking more, and it still wins by a mile.

Thinking is on by default on DeepSeek and reasoning tokens bill as output, which is where the verbosity comes from. It can be turned off. If you do that, you also give up part of the score that put Flash above two flagships in the first place, so measure both ways on your own prompts before deciding.

The practical version of this section: sticker price per token is a bad proxy for cost. Two models with a 27x rate difference finished the same work 22x apart, and two models with a 1.5x rate difference finished 2.4x apart. Token counts vary between models more than prices do.

What Does This Cost You Per Month?

A benchmark suite is not your workload. Here is the same comparison on a concrete team, using the ofox catalog rates from the specs table.

Scenario: 5 developers running a coding agent.

• 40 tasks per developer per day, 21 working days
• Median task: 12,000 input tokens of repo context, 3,000 output tokens
• 40% of input hits the prompt cache

Monthly volume per developer: 10.08M input tokens (4.03M cached, 6.05M uncached) and 2.52M output tokens.

That lands at 36.5x between K3 and Flash, close to the 33.8x the index run showed. Two completely different workloads producing the same ratio is a reasonable sign the ratio is real.

One correction to that table, in Flash’s disfavour. It assumes every model emits 3,000 output tokens per task, and the index run says that is wrong: Flash generated 2.1x the output tokens Qwen did for identical work. Scale Flash’s output line by 2.1x and its team cost goes from $7.82 to about $11.70 a month, which moves the gap against K3 from 36.5x down to roughly 24x.

Still 24x. The verbosity tax is real and it does not come close to closing the gap.

Where this scenario would mislead you:

• Cache hit rate is doing a lot of work here. At 40% hits, DeepSeek’s near-free cached reads barely register because the absolute numbers are small. On a workload with a 200K-token system prompt read hundreds of times a day, that column becomes the entire bill and the gap widens further.
• It prices tokens, not outcomes. If K3 finishes a task that Flash retries three times, the retry cost and the engineer time swamp everything above.
• Rates move. Both DeepSeek builds have changed price and score inside the last quarter. Pull current numbers before you commit.

How Fast Are They?

Qwen 3.7 Max is 5.9x faster than Kimi K3 on output throughput.

Artificial Analysis flags K3 as notably slow and V4 Pro as below average, while Qwen 3.7 Max sits well above average for its price class. There was no speed measurement published for the 0731 Flash build at the time of writing, so that row stays empty rather than borrowing the previous build’s number.

Speed changes different things than score does. A 12-point index gap shows up as tasks the model can or cannot finish. A 5.9x throughput gap shows up as whether a 40-minute agent run becomes a four-hour one. For interactive tooling and long autonomous loops, the second is often the constraint people actually feel.

When Should You Pick Kimi K3?

When the task fails on the other three and the failure costs more than the tokens.

Concrete signals:

• Your evals show the other models failing a class of task rather than doing it slightly worse.
• You need image input alongside frontier reasoning. We tested this through the gateway on 2026-08-03: sending an image raised prompt tokens from 97 to 189 and returned an accurate description, so vision is live and not just listed.
• You need a very large single response. K3’s 1,048,576 max output is 16x Qwen 3.7 Max’s ceiling.
• You want open weights on the top scorer, which K3 is and Qwen 3.7 Max is not.

Signals it is the wrong pick: high request volume, latency-sensitive UX, or any workload where you have not actually measured a quality difference. At 35 tokens per second and $15 per 1M output, K3 punishes both volume and impatience. The Kimi K3 against GPT-5.5 and Opus 4.8 comparison covers how it holds up outside the Chinese-model bracket, and how to use Kimi K3 covers setup.

When Should You Pick DeepSeek V4 Flash?

As the default, until something makes it fail.

It scores second of four, costs least by an order of magnitude, has a 1M context window and a 384K output ceiling, and ships MIT-licensed open weights. The case against it is narrow:

• Text only. No image input, so any pipeline touching screenshots or documents-as-images is out.
• Verbose. 210M output tokens on the index, the most of the four. Budget for reasoning tokens, and cap max_tokens on long loops.
• 7 points below K3. Real, and it will matter for the hardest slice of your work.

If you are currently on DeepSeek V4 Pro, this section is the one to act on. Flash 0731 scores 6 points higher and its index run cost 2.4x less. Any routing rule that sends hard work to Pro and easy work to Flash was written against a build that no longer exists. The V4 Pro against V4 Flash breakdown has the older framing and the update note, and V4 Flash against Gemini 3.6 Flash covers the cross-vendor budget bracket.

When Should You Pick Qwen?

When throughput is the constraint, or when you are already inside the Alibaba ecosystem.

Qwen 3.7 Max’s argument in this group is speed: 206.4 output tokens per second, nearly 4x V4 Pro and nearly 6x K3. For a chat product where users watch tokens appear, or an agent loop measured in wall-clock rather than dollars, that is a real product difference that no index score captures.

The argument against it, on these numbers, is that it is the middle option on capability and near the top on cost, which is an awkward place to be. It scores 4 points below a model that costs 22x less to run the same suite.

And it is no longer Alibaba’s newest. Qwen 3.8 Max shipped 2026-08-03 at $2.00/$6.00, cheaper than 3.7 Max on both sides, with a 131K output ceiling and image input. It is not in this comparison because Artificial Analysis had no page for it at the time of writing, so there is no like-for-like score to place it with. If you are choosing a Qwen today rather than comparing published numbers, 3.8 Max is the one to start from. The 3.7 Plus against 3.7 Max benchmark covers the cheaper sibling, and the 3.7 Max developer guide covers the verbosity tax on long agent sessions.

When Should You Pick None of These?

Three cases where the whole bracket is the wrong shelf:

• You need video input. None of the four take it through the gateway. The Qwen 3.8 Max catalog entry lists text and image only, even though QwenCloud’s own page lists video.
• You need a specific frontier capability these do not have. If your evals are calibrated against Claude or GPT behaviour, a 7-point index difference will not tell you whether the swap is safe. Run your own suite.
• Your bottleneck is not the model. If a retrieval bug or a bad prompt is costing you more than the token bill, none of the numbers in this article will help.

Try All Three via ofox: A/B in 10 Lines

All four models sit in the same catalog, so the comparison is a loop over model IDs rather than four integrations.

Python: run the same prompt through all four

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("OFOX_API_KEY"),
    base_url="https://api.ofox.run/v1",
)

MODELS = [
    "moonshotai/kimi-k3",
    "deepseek/deepseek-v4-flash",
    "deepseek/deepseek-v4-pro",
    "bailian/qwen3.7-max",
]

prompt = "Refactor this module and explain the risk in the migration."

for model in MODELS:
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        max_tokens=2048,
    )
    u = r.usage
    print(f"{model:<32} out={u.completion_tokens:<7} total={u.total_tokens}")

Node: same shape

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.OFOX_API_KEY,
  baseURL: "https://api.ofox.run/v1",
});

for (const model of ["moonshotai/kimi-k3", "deepseek/deepseek-v4-flash", "bailian/qwen3.7-max"]) {
  const r = await client.chat.completions.create({
    model,
    messages: [{ role: "user", content: prompt }],
    max_tokens: 2048,
  });
  console.log(model, r.usage.completion_tokens);
}

Print completion_tokens, not just latency. The whole point of the eval-cost table above is that token counts differ more between these models than prices do, so a per-model output count on your own prompts is worth more than any published index.

One caveat on the catalog. K3’s entry lists /v1/chat/completions only, while the Qwen and DeepSeek entries also list /v1/responses. That field has been wrong in both directions before, so send one real request against the protocol you plan to use rather than planning around the metadata. Current per-model rates are on the model pages for Kimi K3, DeepSeek V4 Flash and Qwen 3.7 Max.

FAQ

Which is the best Chinese model in 2026? On the Artificial Analysis Intelligence Index v4.1 at the 2026-08-03 snapshot, Kimi K3 (max) at 57 is the highest scorer of this group. On score per dollar, DeepSeek V4 Flash 0731 wins by roughly an order of magnitude.

Is Kimi K3 worth the price over DeepSeek V4 Flash? It buys 7 index points for 33.8x the cost of the same eval run. That is worth it only when a specific class of task fails on Flash and succeeds on K3, which you have to measure on your own prompts.

Does DeepSeek V4 Flash have open weights? Yes. The 0731 build is on Hugging Face as deepseek-ai/DeepSeek-V4-Flash-0731 under MIT, in fp8, not gated.

Why is Qwen 3.7 Max in this comparison instead of 3.8 Max? Because 3.8 Max had no Artificial Analysis page when this was written, so there is no score measured the same way as the other three. Qwen 3.7 Max is the newest Qwen with a comparable third-party number.

Which model has the largest output ceiling? Kimi K3, at 1,048,576 tokens. DeepSeek’s two models allow 384,000 and Qwen 3.7 Max allows 64,000.

Do any of these support vision? Kimi K3 does, tested through the gateway on 2026-08-03. Qwen 3.8 Max does. Qwen 3.7 Max and both DeepSeek V4 models are text-only in the catalog.

Is the Artificial Analysis index cost the same as my bill will be? No. It is one specific 9-evaluation suite, so it tells you the ratio between models on a fixed workload, not your absolute spend. Use it to rank options, then measure your own prompts.

Sources Checked for This Refresh

• https://artificialanalysis.ai/models/kimi-k3
• https://artificialanalysis.ai/models/deepseek-v4-flash
• https://artificialanalysis.ai/models/deepseek-v4-pro
• https://artificialanalysis.ai/models/qwen3-7-max
• https://ofox.run/en/models/moonshotai/kimi-k3
• https://ofox.run/en/models/deepseek/deepseek-v4-flash
• https://ofox.run/en/models/bailian/qwen3.7-max
• GET https://api.ofox.run/v1/models (live call, 2026-08-03, for model IDs, prices, context and modality fields)
• Live image-input test against moonshotai/kimi-k3, 2026-08-03: prompt tokens 189 with an image against 97 for the same prompt without one, description accurate