What Kimi, DeepSeek, Qwen and Doubao actually cost from Shenzhen

At two in the afternoon Shenzhen time, my DeepSeek bill doubles. Peak billing runs 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which in Beijing time is 09:00 to 12:00 and 14:00 to 18:00, and if you lay those two blocks on a calendar you have drawn the Chinese office day with the lunch break cut neatly out of the middle.

Off-peak is exactly half of peak on every line: cache hit, cache miss and output. A team in New York running deepseek-v4-pro through their own working day sits entirely in off-peak and pays $0.66 per million input tokens on a cache miss and $1.98 per million output. Doing identical work at my desk, in the hours I am actually awake for it, I pay $1.32 and $3.96.

Same weights. Different clock.

I have moved most of my daily engineering work onto Chinese models over the past several months. Not to benchmark them: to use them. Code review, log triage, spec drafting, translating supplier documentation out of Chinese and a great deal of tool-calling inside agent loops. What follows is what the pricing pages, the licence files and the latency traces look like from inside the mainland, as of late August 2026.

The four price sheets, read properly

DeepSeek's current line is three models: deepseek-v4-pro, deepseek-v4-flash and an experimental deepseek-v4-flash-vision-exp. All three carry a one million token context window with 384K maximum output. Off-peak, V4-Pro is $0.022 per million tokens on a cache hit, $0.66 on a miss and $1.98 output. V4-Flash is a third of that at $0.007, $0.22 and $0.66. Peak doubles every figure, which is the single most important sentence on that page and the one most people skim past.

Moonshot sits at the far end of the same market. Kimi K3 is $3.00 per million input tokens on a cache miss, $0.30 on a hit and $15.00 output, across a context window of 1,048,576 tokens. Call it five times DeepSeek's input price and roughly eight times its output. The K2 family is where Moonshot's value lives, with K2.7 Code at $0.95 input, $0.19 cached and $4.00 output over 256K, and a batch API that bills at 60% of list. Batch does not cover K3.

Qwen3.8-Max arrived on 3 August 2026 at $2.00 input and $6.00 output per million, one million context, 131,072 maximum output and cache reads at $0.25. The rate is flat across the whole window, which matters more than it reads: several competitors step the price up past 128K or 256K and Alibaba does not, so long-context work prices predictably.

Doubao is the one you cannot honestly quote in dollars, because it is not sold in dollars. Doubao Seed 2.1 Pro is ¥6 per million input tokens and ¥30 per million output on Volcengine, which converts to roughly $0.83 and $4.14. Turbo is half of that.

I was in Lagos in the spring, walking a small team through how this maths falls out. They did a sum in their heads that I had not bothered to do, in naira. Their engineers are paid in naira and the API bill arrives in dollars, so the output-token price is not a line item to them, it is the architecture. They went straight to V4-Flash and cached aggressively, and they were right. Most teams I talk to here have never had to think that hard about the cheap tier.

Every foreign team currently paying half price for off-peak inference is arbitraging someone else's lunch break.

The account wall is higher than the technical wall

Volcengine is the hardest of the four to get into and almost nobody writes this down clearly. Registration expects a +86 mobile number for SMS verification. Individual real-name verification expects a Chinese national ID. Funding the account means Alipay or WeChat Pay backed by a mainland bank card. I have watched capable engineers spend a fortnight on this and quietly give up. Doubao's pricing is excellent and largely theoretical if you are not resident here.

Alibaba built the opposite thing on purpose. Model Studio runs as separate regional deployments, including Beijing, Singapore, Tokyo, Frankfurt and Virginia, and they are genuinely separate: different base URLs, different model availability, different pricing and API keys that do not work across regions. That last detail catches people. A key minted in Singapore returns an authentication error against Beijing, and the error message does not explain why.

DeepSeek and Moonshot are the straightforward ones. An email address and a card gets you running from anywhere. Moonshot's platform.moonshot.ai now redirects to platform.kimi.ai, and the page to read before you write a line of code is the rate-limit table rather than the pricing table. Limits are tiered by cumulative top-up and the entry tier is deliberately tiny, a single concurrent request, which will strangle any agent loop on its first run. Top up past it on day one. It costs less than lunch in Futian and it saves you an afternoon debugging 429s that are not your fault.

Four vendors, four different meanings of the word open

DeepSeek is the only one of the four whose flagship is genuinely unencumbered. V4-Pro is published to Hugging Face and ModelScope under an MIT licence at 1.6 trillion total parameters and 49 billion active per token, in FP8 mixed and FP4 plus FP8 mixed formats. V4-Flash went the same route at 284 billion total and 13 billion active. MIT means MIT: no revenue threshold, no branding clause, no field-of-use carve-out.

Kimi K3 is widely reported as modified MIT and that reporting is wrong. The LICENSE file in the repository is a bespoke document titled Kimi K3 License, and the Hugging Face metadata records it as license: other with license_name kimi-k3. The terms bring in a separate agreement with Moonshot for model-as-a-service businesses above a revenue threshold, and a branding obligation above a scale threshold. Most teams will never approach either line. Your counsel will still want to read the file, and they are right to.

Alibaba split the difference inside a single week in August. The Max-class Qwen3.8 weights shipped under a custom Qwen licence. Two days later the 27B model landed under Apache 2.0. The old shorthand that Qwen is the Apache vendor now holds only for the small one, and that distinction is exactly the kind of thing that gets flattened in a comparison table written by someone reading press releases.

Doubao publishes nothing. API only.

One practical note from this side of the firewall. DeepSeek and Alibaba both mirror to ModelScope, and on a Shenzhen connection ModelScope saturates the line while Hugging Face crawls or stalls without a proxy. Pulling several hundred gigabytes of V4-Pro weights is a completely different afternoon depending on which mirror you point at. Teams outside China have the exact inverse problem and almost never think about it, because Hugging Face simply works for them.

Latency, and why I stopped trusting a single reading

Artificial Analysis currently places Kimi K3 at maximum effort first out of the models it tracks on its Intelligence Index, scoring 60, with DeepSeek V4-Pro-0813 at maximum reasoning effort second on 53. Read only that line and you would pick K3 and move on.

Serving numbers invert the decision. On the same comparison, V4-Pro produces 68 output tokens per second against K3's 36 and reaches first token in 1.65 seconds against 6.51. Blended cost per million comes out at $0.69 versus $2.31. Seven index points, against a model that is twice as fast, four times quicker to start speaking and under a third of the price.

Now the honest part. I pulled that same comparison twice inside an hour and got materially different figures: 80.3 tokens per second and 2.79 seconds to first token on the first read, then 68 and 6.51 on the second. Neither reading is wrong. Both are real measurements of a shared endpoint carrying whatever load it happened to be carrying. If your design has a first-token budget, measure it yourself, from your own egress, at the hour you will actually run it, for a week.

DeepSeek publishes concurrency limits openly at 500 on Pro and 2,500 on Flash. Moonshot gates concurrency behind cumulative spend instead. That structural difference shapes how you build a fan-out job long before anyone argues about model quality.

The benchmark OpenAI stopped reporting

In February 2026 OpenAI's evaluations team announced it would stop reporting SWE-bench Verified, and the reasoning is worth reading in full if you have ever quoted a Verified score in a deck. A manual audit of 138 difficult tasks found that in 59.4% of them the test design or the problem statement itself was flawed. A further slice used tests narrow enough to reject correct solutions that took a different route to the same result.

Contamination was the other half of it. A frontier model solved 31 tasks the harness had classified as nearly impossible, and in at least one case its chain of thought already contained release-note text describing the change it was supposedly reasoning its way to. OpenAI now reports the public split of SWE-bench Pro, 1,865 multilingual tasks, and recommends other labs do the same.

None of this is a Chinese-model problem specifically. It is a benchmark problem and every lab is inside it. What it does mean is that a seven-point index gap between two frontier models is smaller than the noise in your own workload, and that a coding-benchmark headline from any vendor in 2026 tells you less than one afternoon spent running the model against your own repository with your own tests.

Build the eval out of your own closed pull requests. It takes a day. It is the only number that survives contact with your codebase.

Chinese in, English out

The place these models are simply better is Chinese input, and it stays underrated because most people writing comparisons never have any. Earlier this year I spent a morning in a workshop in Longhua with a supplier's process document: fifty-odd pages, GB/T standard references throughout, shop-floor abbreviations that appear in no dictionary and a third of the tables scanned at an angle. Qwen and Doubao read it correctly, abbreviations included. The Western models I tried produced fluent, confident, subtly wrong English, which is the worst failure mode available because it survives review.

English output runs the other way. Qwen and Doubao write English that parses cleanly and sits half a step wrong in register: over-formal code comments, connective tissue that reads translated, a fondness for structure where one sentence would do. Kimi K3 is the strongest English writer of the four by a clear margin, and that is part of what $15 per million output tokens buys. DeepSeek V4-Pro is close behind at a fraction of the cost. For anything customer-facing I still read every line.

Tool calling is where the distance between a demo and production actually lives. Single-call accuracy on all four is high enough that a five-minute test tells you nothing useful. What matters is consistency down a chain, and the arithmetic there is unforgiving: a step that succeeds 90% of the time succeeds eight times running only 43% of the time, and real agent loops are twenty and thirty steps deep. In my own runs DeepSeek V4-Pro is the most consistent at emitting well-formed arguments against a long schema, Kimi K3 recovers best when a tool hands back an error and Qwen drifts fastest once a conversation passes roughly forty turns. Those are my workloads, not a benchmark. Run yours.

Everyone tests the first call. Production is the twenty-eighth.

The strongest case against all of this

Here is the argument I cannot beat, stated at full strength rather than in a version I could knock over. Benchmark parity is not workflow parity. Your prompt library, your cache strategy, your evaluation harness, your observability stack and your fine-tunes are all shaped around one provider's specific behaviour. Migrating is weeks of engineering for a saving that, for most teams, is a modest line in a budget dominated by salaries. Cheaper tokens do not pay for a quarter of rework.

Procurement is harder still. DeepSeek is restricted on government devices across a growing list of jurisdictions including Australia, Taiwan, Italy and the Netherlands, at multiple US federal agencies and in a number of US states, with legislation moving through Congress to formalise the federal position. Italy's data protection authority acted against the DeepSeek corporate entities directly. If you sell to any of those buyers, the technical merits are not the question anyone is asking you.

Data residency has a partial technical answer that most critics skip. Alibaba's Frankfurt and Singapore regions are separate deployments with separate keys and separate pricing, and a European team can terminate traffic in Frankfurt without a packet crossing into the mainland. That answers residency. It does not answer the corporate entity, the sub-processor list or the government-access question, and a procurement officer who understands the difference will not accept an endpoint as an answer about ownership.

Which leaves exactly one configuration that clears a strict Western procurement process: weights on hardware you control, under a licence your counsel will sign without a call. That is DeepSeek V4-Flash under MIT, or the 27B Qwen under Apache 2.0. Not the API. Not the flagship. The consequence is quietly funny. The most heavily regulated buyers end up on the smallest models, and the smallest models are the ones nobody writes leaderboard threads about.

The clock is set to Beijing

Go back to the peak window. DeepSeek did not pull 01:00 to 04:00 and 06:00 to 10:00 UTC out of a hat, and it is not purely a GPU utilisation curve either. Those hours are a Chinese working day with a lunch break in the middle, expressed in a timezone that is not China's.

Every foreign team currently paying half price for off-peak inference is arbitraging someone else's lunch break. The discount exists because they are not the customer the schedule was drawn for.

That is a comfortable position to occupy. It is also a fairly precise measurement of how little the mainland market currently cares what time it is in San Francisco. The rates will change the week that stops being true.

Sources

DeepSeek API models and pricing (official documentation) : https://api-docs.deepseek.com/quick_start/pricing

DeepSeek-V4-Pro model card and MIT licence, Hugging Face : https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

Kimi-K3 model card and licence metadata, Hugging Face : https://huggingface.co/moonshotai/Kimi-K3

Kimi K3 API pricing and context window, OpenRouter : https://openrouter.ai/moonshotai/kimi-k3

Qwen3.8-Max API pricing and context window, OpenRouter : https://openrouter.ai/qwen/qwen3.8-max

Doubao Seed 2.1 Pro pricing on Volcengine, LLMReference : https://www.llmreference.com/model/seed-2.1-pro/volcengine

Kimi K3 vs DeepSeek V4 Pro: intelligence, speed and price, Artificial Analysis : https://artificialanalysis.ai/models/comparisons/kimi-k3-vs-deepseek-v4-pro

Why we no longer evaluate SWE-bench Verified, OpenAI : https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/