Intelligent Pricing: DeepSeek, Kimi, and the Varied Choices of Chinese AI Companies

By | Zhu Yingli

Edit丨Zhao Lei

In May of this year, Xiaomi announced a permanent price cut for the MiMo-V2.5 series API, with discounts up to 99%. In June, Tencent Cloud lowered the prices of the DeepSeek-V4 series on its platform, with discounts up to 97.5%. In August, DeepSeek adjusted the V4 series pricing, introducing peak‑off‑peak differentials. In September, with the release of new models, prices dropped again.

Such news is becoming more common. Model companies are eager to tout their prices. How much they’ve dropped, a few cents per million tokens, several times cheaper than competitors—these are easy‑to‑share figures. It appears that large models are turning into a standard commodity, with token offerings from different companies being largely similar; the lower the price, the stronger the competitiveness.

In reality, a token is merely a unified billing unit; the cost of tokens varies across providers, and the intelligence value per token also differs. DeepSeek emphasizes “reasonable profit,” Kimi focuses on “ultimate performance,” and other model companies have their own priorities—an intriguing divergence has been gradually forming over the past six months.

Pricing a model has become a tool for companies to screen customers, allocate compute resources, improve profitability, and even shape external technical narratives.

The essence of price is, in one word: value,” says Herman Simon, author of *Pricing to Win*.

But for large‑model products, the technology evolves monthly while applications have not been fully validated, making “value” hard to quantify and pricing methods difficult to compare with other mature goods.

Based on publicly available information and statements from commercial staff at several large‑model vendors, we find that there are basically only two pricing approaches for large‑model products today: cost‑plus pricing and market‑positioning pricing. Some companies lean more toward the cost side, such as DeepSeek and Xiaomi Mimo, while most other vendors focus more on market positioning.

Cost‑plus pricing is relatively simple: the company calculates how much it costs to continuously provide the service, adds the desired profit, and arrives at a price. This cost mainly includes inference compute, networking, storage, and operations, and does not account for upfront model R&D expenses or other operating costs such as sales and salaries.

DeepSeek is the most typical example. In May this year, Liang Wenfeng explained DeepSeek’s pricing principle at an investor briefing: the API business aims to recoup equipment investment within ten months. In his view, this is an acceptable return, and there is no need to raise prices to maximize revenue or profit.

According to his calculation, recovering costs in ten months with an approximate five‑year depreciation period corresponds to about a six‑times profit margin. A simplified calculation: assuming cost only considers depreciation, the required gross margin would be about 83.3%. According to a report cited by The Information from insiders, DeepSeek’s API business gross margin for the first seven months of this year was 82.9%, close to this estimate.

This approach is also reflected in subsequent price adjustments. When V4.1 Flash was released in September, the company said the new architecture reduced the KV Cache’s demand for high‑bandwidth VRAM to one‑quarter of the previous generation and its demand for SSD to one‑eighth. Afterwards, the prices for cache hits and output tokens also dropped.

Reducing redundant computation, compressing caches, and increasing throughput can all create room for price cuts. But cost isn’t the only factor; to ensure compute and hardware are fully utilized, DeepSeek also adjusts prices based on demand.

For example, in August DeepSeek introduced peak‑off‑peak pricing, with off‑peak calls costing only half of peak‑time rates. After the model price cut in September, the peak‑off‑peak differential remained. This pricing strategy eases peak‑time congestion while allowing idle equipment to earn revenue.

A commercialization employee at a model company described DeepSeek’s pricing as having a “user‑centric” vibe: “They defined a special product category space and then drove the changes within that space.”

Most other model vendors don’t price this way. Beyond inference costs, they primarily consider market positioning—finding their spot within the performance‑price reference framework of numerous models.

According to employees at large model manufacturers, when pricing each model they always think about which price tier to target. “Essentially it’s about model positioning; everyone works toward that positioning.”

Before a new model is released, the commercialization team usually works with the Infra team to calculate inference costs under real workloads, then references third‑party benchmarks, internal beta customer feedback, and business goals. The price is typically decided by multiple departments together.

Take Kimi’s K3 model as an example: it’s a 2.8 T parameter model with about 104 B activation parameters per token, and its operating cost is not low.

Nvidia distributor Openzeka tested K3 on an 8‑GPU B300 server. According to the test, the machine can process roughly 4.05 million input tokens and 4.05 million output tokens per hour. If we roughly calculate using Runpod’s B300 single‑GPU price of $8.99 per hour, handling “one million input tokens plus one million output tokens” would cost about $17.7 in compute rental alone.

This will inevitably differ from the actual cost on the “dark side of the month,” but it already shows that K3 is unlikely to become a cheap model.

With a rough cost range in hand, the next step is to find a benchmark product.

On July 17, Artificial Analysis calculated that K3’s overall capability is close to Claude Opus 4.8 and GPT‑5.5. At that time, the input prices for the two models were $5 and $4 per million tokens, and the output prices were $25 and $20 respectively.

Kimi ultimately set K3’s price at $3 for input and $15 for output.

This figure also reflects the company’s strategic choice. After K3’s release, the B‑side head of Moon’s Dark Side publicly said they do not want Chinese models to be constantly labeled as low‑price. When capability reaches a cutting‑edge level, it should be sold at a reasonable price. Compared with earlier models, K3 aims to retain a larger profit margin and try to establish a premium for a flagship model.

Customer testing in real‑world scenarios also influences pricing. An engineer from a model vendor said that people are increasingly focusing on “per‑token intelligence” rather than pure cost‑performance. If a small‑scale customer test shows that a task can be completed with fewer tokens, the unit price per token will rise; otherwise, it will be adjusted downward.

Artificial Analysis once had K3 and GLM-5.3 perform the same task. K3 consumed about 160 million tokens, while GLM-5.3 used roughly 200 million. K3’s token price is higher, so the total cost for both to complete the task is comparable.

Costs, competitors, and test results determine the rough price. There will also be some finer trade‑offs. Kimi, for example, decided that for the same model the price in the Chinese market should be lower than overseas, “even if it’s just a tenth of a yuan.”

Channels also affect pricing. Besides the vendor’s direct API, models can be resold or hosted by cloud platforms, routed through aggregation platforms, or deployed by customers themselves. The vendor’s public pricing is relatively fixed. Other channels must consider upstream procurement costs, hardware prices, deployment efficiency, and operational expenses. Within licensing limits, cloud providers and compute service vendors can deploy open‑source models themselves and quote based on their own costs and efficiency, potentially squeezing the vendor’s revenue.

Liang Wenfeng mentioned this at an investor Q&A session. He judged that if the vendor demands an extremely high margin, even a less efficient third‑party can win customers with lower prices. DeepSeek, charging at a rate that recoups hardware investment in ten months, has already pushed the price to a level that third parties find hard to match, allowing the official API to retain a price advantage while still making a profit.

According to our research, the vast majority of secondary compute providers cannot optimize the DeepSeek model to the level claimed by the official specifications.

The birth process of a model’s price is convergent, but different vendors’ approaches can shape distinct pricing outcomes.

Kimi K3 and Alibaba Qwen Max are two large models of comparable size. K3 has a total of 2.8 T parameters, with about 104.2 B active parameters per token; Qwen Max has 2.4 T total parameters and 95 B active parameters. Both are in the same magnitude, yet their prices differ significantly. Qwen Max’s input token price is 60% of Kimi’s, and its output price is less than half of Kimi’s.

Alibaba Cloud’s sales team emphasizes the “cost‑performance” of the Qwen model, while Kimi’s staff believe that for high‑performance models, “performance” is more important than price.

In Kimi’s strategic positioning, their more important goal is to explore the “intelligence ceiling”—to enable the model to handle more complex, longer‑running tasks and find customers willing to pay for such capabilities. “Only when your performance reaches a certain level does a price‑cut strategy make sense. Essentially, customers won’t switch to someone else just because you’re a few cents more expensive,” said a source close to Kimi.

He believes that ultra‑high performance is certainly not the most common market demand, but someone in China needs to do this. “Although from a profit perspective it may not be the most lucrative,” he adds, “the flagship model should charge for stronger capabilities; there’s no need to chase the widest usage from the start.”

Other companies may not agree. An algorithm engineer at a model company thinks Kimi’s high price will limit usage scale, and that the top priority for large‑model companies right now should be real‑world usage volume, because only that creates a data flywheel. “If you set a price comparable to the U.S., you’re basically saying you don’t want to sell,” he said.

However, according to a senior AI product manager familiar with various models, Kimi has actually found a unique survival path outside the domestic API market: going overseas to serve as a post‑training “foundation” for Silicon Valley application layers.

For example, in March this year, Cursor launched Composer 2, which continues training on Kimi K2.5. In July, Moonlight Darkside and Together AI announced a partnership, with the latter offering K3 via infrastructure deployed in the United States and allowing customers to perform post‑training with their own data. In August, legal AI company Harvey released a research preview of the Tenet model based on K3.

The Kimi team also erected a huge billboard in the San Francisco Bay Area, proactively reaching out to more overseas application companies and alleviating the commercialization pressure of large models through high‑value technical collaborations.

For cloud providers such as Alibaba Cloud and Volcano Engine, model APIs are both a business and an entry point for customers into AI cloud services. Once enterprises integrate models into production systems, they typically also need compute power, development tools, data processing, deployment, and long‑term operations. Cloud providers can acquire customers with inexpensive models and then monetize through subsequent services.

Therefore, the price a company sets for a token largely depends on where it intends to make a profit.

The different choices became more evident in this year’s price adjustments across companies.

DeepSeek has consistently emphasized “restraint.” Liang Wenfeng once used a hypothetical at an investor briefing to explain his thinking: “If AI were to capture 20% of global GDP, and my goal were to take 5% of that, the math would work out, just like OpenAI; but it would inevitably be beaten by someone who is satisfied with only 1%, and that 1% would be beaten by someone aiming for just 0.1%.”

He doesn’t want to set profits at a level that gives competitors enough incentive to fight for it. Recovering equipment investment within ten months is the line DeepSeek has chosen. Profits beyond that line may not be worthwhile. By keeping prices low enough to ensure reasonable profit while reducing third‑party deployment competition, a balance is struck between affordability for users, recouping equipment costs, and allocating limited compute.

Improvements in technical efficiency create the conditions for this balance. When DeepSeek released V4.1 Flash in September, it explicitly linked the API price cut to cost reductions from the new architecture. Yet when compute is tight, it will also curb demand through pricing.

Zhipu has also been adjusting prices frequently this year, but mainly raising them.

As the open platform and API business become the main source of Zhipu’s revenue, API pricing has a more direct impact on Zhipu’s operations. In the first half of 2026, revenue from Zhipu’s open platform and API business reached 825 million yuan, accounting for 86.5% of the company’s total revenue; the gross margin of this business rose from –0.4% in the first half of 2025 to 24.6%. Zhipu attributed the improvement in its financial report to scale effects, price hikes, and reduced inference costs.

Price increases have become a way for this listed company to improve profitability.

In the past, relying on the academic influence of Tsinghua-affiliated scientists such as Tang Jie, Zhipu built a good reputation in the tech community and among government and enterprise clients, but the development of its commercialization team lagged noticeably.

In February this year, Zhipu apologized for a revision of the GLM Coding Plan. When the new model was released, traffic exceeded expectations, and because scaling was not timely, some users could not obtain access promptly or experienced throttling. The company admitted that the rules were opaque and the upgrade mechanism was crudely designed, and offered affected users a refund option.

A recent employee who moved from a big tech company to Zhipu to handle regional commercialization discovered that the East China commercialization team at Zhipu is extremely small; she had to work nonstop on her own to support major exhibitions right after joining. When visiting a client, a business representative from a leading automaker complained that it was impossible to find a contact person at Zhipu for cooperation and they could only call the 400 number on the official website; another client had to go through the Tsinghua alumni network to connect with Zhipu’s technical staff just to get a deal moving.

The model’s capabilities have already attracted customers. For Zhipu, the current challenge is how to turn this technical advantage into sufficient revenue and profit through commercialization, enough to fund the next round of R&D and meet the market’s profit expectations.

In contrast, cloud vendors’ model pricing structures are much more stable.

A product manager of a cloud vendor’s MaaS platform said that large‑scale customers value stability the most; a long‑term stable list price creates a better customer experience, and even when underlying optimizations reduce costs, they are translated into private “differentiated commercial discounts” for clients of different scales.

Alibaba Cloud and Volcano Engine both already have mature sales systems, operation teams, and a massive enterprise customer base. For companies like these, price itself is a negotiation tool.

However, both Alibaba Cloud and Volcano Engine’s Maas platforms recently launched a ¥9.9 acquisition promotion, and cloud vendors familiar with enterprise business are also starting to use almost loss‑making low‑price packages to compete for individual developers. For them, how to truly attract a broader range of individual developer users and increase the influence of self‑operated models is a new test.

The differentiation in large‑model pricing continues. As model capabilities improve and inference costs decline, users are handing increasingly complex and time‑consuming tasks to AI; yet beyond measuring token consumption costs, there is still no better way to gauge how much value AI creates.

Image source: “Big but Not Falling”

– FIN –