China's large-model price war ends, but the cost war has only just begun

China's large-model industry fought a price war that lasted more than two years. In May 2024, DeepSeek-V2 pushed the input price down to one yuan per million tokens, and ByteDance, Alibaba, and Baidu took turns following suit within a week. Prices kept falling and did not stop afterward: in February 2025, Baidu announced that its chatbot product Ernie Bot would become fully free starting in April, and Ernie 4.5, released in March, pushed API prices down to one percent of GPT-4.5's.

According to monitoring by the China Academy of Information and Communications Technology, by the second half of 2025, the normalized average price of domestic large-model token services settled at 6.8 yuan. It should be noted that this average price is a normalized statistic based on the quoted prices at the time each model was released, not the average price on users' actual bills.

This October, the situation seemed to reach a resolution. In mid-to-late October, StepFun, DeepSeek, MiniMax, iFlytek, and Moonshot AI released new models in succession within ten days. Across the five launches, no new price war emerged, yet changes in customer bills were greater than in the past two years. What happened in between must be explained starting from how the price war came to an end.

How the price war came to an end

After the average price fell to the low point of 6.8 yuan, the large-model API market did not continue to fall in unison; instead, it split by model tier. In the first half of this year, price cuts and price increases occurred simultaneously. In May, Xiaomi implemented a permanent price cut on the MiMo-V2.5 series, with a maximum reduction of 99%. In June, several general-purpose chat models also saw price cuts exceeding 90%. But under the same methodology, the normalized average price rose to 13.46 yuan, an increase of about 97.9%.

The price gap was pushed up by flagship models. Zhipu raised prices three times during the year, and CEO Zhang Peng said this was to return to normal commercial value; DeepSeek introduced peak and off-peak pricing for the V4 series, with output prices rising from 6 yuan to 27 yuan during peak hours.

In July, Moonshot AI released K3, raising output to 100 yuan in one step and making it the most expensive domestic large model. This round of price increases did not trigger any lasting follow-up. According to media reports, its ARR instead rose from $300 million in June to nearly $1 billion by September. Low prices in the inclusive tier remain, while the flagship tier is collectively moving upward. This means the era of winning customers with a single low price is over.

The October that has just passed made this change concrete with five new models. From October 15 to 24, StepFun, DeepSeek, MiniMax, iFlytek, and Moonshot AI launched in succession. The product form was still model plus API, and among the five, only Kimi lowered its quoted price; the other four kept pricing flat with their respective previous generations.

DeepSeek's per-task cost fell from $0.40 to $0.26, MiniMax cut Agent costs by 70%, iFlytek competed on keeping data local, and StepFun pushed per-task cost down to one-eighth of Claude Opus 5's. The only one to cut prices, Kimi, saw almost no change in its per-task cost in Artificial Analysis's actual tests. The unit of measurement each company competes on has shifted from the model's list price to the money a customer actually spends to complete a task.

The price war is over, but the real cost war that lands in business has only just begun.

DeepSeek: prices unchanged, bills halved

On October 17, DeepSeek released V4.1 Flash-1017. The official release notes said this update was a Pareto improvement in performance and cost. Official data shows that its score on the software engineering benchmark DeepSWE v1.1 surpassed Claude Opus 5 and GPT-6 Astra. However, on several other tests such as Terminal-Bench 4.0, Opus 5 still leads. DeepSeek's strength has never been comprehensive superiority, but approaching it at very low cost.

What is truly interesting about this update is the optimization for a specific problem.

The September version of V4.1 Flash had an obvious shortcoming: it was verbose. In Artificial Analysis's tests, it needed to output about 89,000 tokens on average to complete an intelligence index task, 62% more than the previous generation V4 Pro, and part of its price advantage was eaten up by the tokens it wrote.

The 1017 version is precisely about "saying less nonsense." Measured at Artificial Analysis's highest reasoning tier, per-task cost fell from $0.40 in the September version to $0.26 in the 1017 version. On the overall ledger, K3.1 Pro costs about $2.11 to run an intelligence index task, while DeepSeek's 1017 version costs only $0.26, an eightfold difference.

It no longer relies only on being the cheapest; instead, with unit prices unchanged, it uses engineering optimization to lower customers' actual spending.

This playbook was doubled down on after the results were delivered. On October 20, DeepSeek open-sourced its internal inference framework DeepInfer, compatible with Nvidia and Huawei Ascend. Based on this framework, a single cluster has achieved a service throughput of 155 trillion tokens per day, generating $38 per card per hour at off-peak prices.

A week later, on the evening of October 31, the DeepSeek Harness team open-sourced the Super-agent prototype architecture and a 10,000-word design document: the always-on Router layer only coordinates and never does the work itself, while the Worker layer executes tasks in parallel in independent contexts. The underlying layer uses event-driven design and separation of storage and compute to reduce waiting and redundant data movement, and an external supervisor process monitors the execution process; whichever task gets stuck or fails is automatically restarted or taken over, with no need for a human to step in.

In preliminary experiments, half of the tasks were completed within 12.4 minutes and 95% within 15.7 minutes, and batching at the routing layer cut the time amortized per single decision from 6.4 seconds to 1.6 seconds. This system is not yet a product, and the experiments are based on synthetic workloads, but this open-sourcing points to one item in the cost structure that has not yet been compressed. After token costs are pushed down, the remaining big chunk in Agent tasks is people: tasks require people to issue instructions, people to watch progress, and people to rearrange things after errors. The architectural goal of Super-agent is to reduce dependence on human intervention during Agent execution and free people from the waiting and handling involved in every step.

Compressing costs addresses the top half of the income statement; the revenue side must also be able to scale up for this business to hold. According to media reports, DeepSeek's annualized revenue has exceeded $1 billion, up from less than $500 million just a few months ago. But the cost is equally obvious. On Vercel AI Gateway, a model aggregation platform commonly used by developers, DeepSeek's models accounted for 58% of all tokens called through the platform in September, but the fees generated by those calls accounted for only 6.6% of the platform's call revenue. Call volume is enormous, yet revenue contribution per token is very low.

Reflected in the investment and financing market, on October 30, details of DeepSeek's second funding round were disclosed: 105.5 billion yuan, with a post-money valuation of 615.5 billion yuan, nearly 30 institutions participating, the state-owned share rising from 2% in the first round to 27%, and compute-industry-chain companies such as ChangXin Memory Technologies and SMIC entering the list. Combined with the STAR Market preparation launched in late August (with CITIC Securities as the tutoring broker and a plan to file as early as the end of the year), the company's moves form a complete chain: open-source models to grow call volume, an open-source inference framework to push down service costs, and tens of billions in financing to guarantee compute investment. All three links serve the same goal: to keep lowering the inference cost per million tokens.

MiniMax: cutting costs, speeding up

MiniMax's drive to cut costs is more direct than DeepSeek's, because it is a listed company and its income statement is publicly disclosed every quarter.

On October 19, MiniMax released M3.1 Flash. Its parameter scale is the same as the previous-generation flagship M3, and the changes are concentrated in the model structure: several full-attention layers retained in M3 were replaced with sparse attention layers in the new model.

MiniMax says the new model's online inference speed has increased nearly threefold, and the comprehensive usage cost for Agent-type tasks is 70% lower than the previous generation. API pricing is 1 yuan per million input tokens and 4 yuan per million output tokens. In official evaluations, it scored 61.7 on the OSWorld V2 computer-operation test, the highest among domestic models, and tied for first with Opus 5 on BankerToolBench.

Worth noting is the choice of test items MiniMax disclosed: besides programming, it focused on enterprise process automation across SaaS systems, computer operation, and BankerToolBench for investment banking work.

The value of speed gets amplified in actual tasks. An industry media outlet ran a comparative hands-on test: for the same development task, MiniMax Code plus M3.1 Flash took about 55 minutes, while another setup using OpenCode with DeepSeek V4.1 Flash took about 121 minutes, with model fees on both sides around 6.5 yuan. Same spending, but the delivery time differed by more than twofold.

For MiniMax, cutting costs is not just a bonus. It listed on the Hong Kong Stock Exchange in January this year at an issue price of HK$165, and its first interim report after listing showed revenue of $117 million in the first half of this year, up 283% year on year, with the open platform and enterprise services already accounting for more than 60%. But gross margin was only 17.8%, adjusted net loss was $293 million, and R&D investment was $297 million. The high revenue growth has not translated into improved profitability, and losses are still expanding along with R&D investment.

For MiniMax, cutting costs is the necessary path to improving profit.

iFlytek: not competing on price, selling "data stays local"

The previous three companies all compete in the same developer market, comparing whose math works out better. But the next two have chosen another path.

On October 22, iFlytek released Spark X2.6. The keyword for this release is "device-cloud integration."

Spark X2.6 includes a cloud large model and an open-source on-device model, X2.6-4B. iFlytek's plan is: in the future, 60% of applications will be completed locally, 30% through device-cloud collaboration, and only 10% will need to be handed to the cloud large model for processing.

This set of proportions comes from iFlytek's customer structure. iFlytek's large models first landed in education, healthcare, and law: AI lesson preparation and teaching research, medical imaging and health management, and a legal assistant covering the entire case-handling process. When institutions in these fields choose a model, whether the data can stay in their own hands is often more important than the unit price.

Spark X2.6 was trained entirely on domestically produced computing power throughout. "Only by truly training your own model on domestic computing power can you be truly autonomous and controllable," said Chairman Liu Qingfeng.

This route has its engineering costs. Domestic chips have limitations in memory and bandwidth, and iFlytek made dedicated optimizations for this, improving training efficiency for ultra-long contexts by about 4 times and inference efficiency by 4.5 times.

The gains and costs of this path are reflected simultaneously in iFlytek's financial reports. In the third quarter of this year, iFlytek's revenue was 6.564 billion yuan, up 7.99% year on year. Among this, the smart education and smart healthcare businesses both achieved growth of more than 20%, while G-end business declined slightly due to a proactive contraction of the smart city business. The cost is R&D investment: more than 1.7 billion yuan in the third quarter, accounting for more than a quarter of revenue, and it still posted a loss of 126 million yuan after deducting non-recurring gains and losses.

There were corresponding organizational moves as well. In June this year, iFlytek established a wholly owned subsidiary, Ciyuan Xinghuo, specifically responsible for the R&D and commercialization of general large models, and explicitly proposed to achieve independent accounting and clearly distinguish investment from returns.

This adjustment means that, within iFlytek's business landscape, the large model is moving from a cost center to an independent profit center, and its input-output will for the first time be subject to separate assessment.

StepFun: open source in exchange for ecosystem, betting on devices

The fourth report card comes from StepFun, and its difference from the previous three lies not in model capability but in the monetization path.

DeepSeek, MiniMax, and Moonshot AI all derive their main revenue from APIs, charging developers by token, competing over whose calls are cheaper and faster. iFlytek sells industry solutions to schools, hospitals, and legal institutions. StepFun also offers APIs, but it does not treat the developer market as its main battlefield. Its commercial vision is to put models into terminal devices and generate revenue from on-device deployment in phones and car cockpits. The shareholder list mentioned later was assembled for this path.

Step 5 Preview is StepFun's new-generation flagship foundation model, launched on September 20 and officially open-sourced on October 15. It has 600 billion total parameters, activates about 27 billion per token, supports a 1-million context, and targets scenarios such as AI programming, software engineering, professional knowledge work, and finance.

Compared with flagships that routinely have trillions of parameters, this scale is restrained. But what StepFun cares about is not scale but cost-effectiveness: according to official estimates, at a comparable level of intelligence, its per-task cost is only one-eighth that of Claude Opus 5. Soon after launch, it entered the global top three open-source models in Artificial Analysis, with a current intelligence index of 44 points, and feedback from trials in the developer community also concentrated on the same direction: capability is in the same tier as GLM-5.3 and Kimi K3, at a lower price. It should be noted that developer reputation here is not a source of revenue but an ecosystem signal; when device manufacturers decide whether to install a model on a device, it is precisely these public benchmarks and community feedback that they reference.

When the weights were open-sourced on October 15, Huawei Ascend, Biren Technology, MetaX, and Moore Threads simultaneously announced first-day support.

As soon as a model is open-sourced, it can be deployed directly on multiple types of hardware, which lowers the barrier to use for downstream manufacturers. In January this year, StepFun completed a B+ round of more than 5 billion yuan, with Yin Qi, co-founder of Megvii and chairman of Qianli Technology, appointed as chairman. In May, StepFun was reported to be completing a new round of nearly 2.5 billion US dollars.

This round of financing has a strong industrial character. In addition to financial investors, it includes phone ODM manufacturers Huaqin and Longcheer, image sensor maker OmniVision, and terminal industry chain companies such as ZTE. Tencent Cloud also signed a strategic cooperation agreement with StepFun, and the two sides will jointly build AI Agents in car cockpits.

The supporting environment for its device route has also improved over the past week. On October 31, a domestic AI chip company released the industry's first solution that runs locally on phones, capable of running a 13-billion-parameter model without an internet connection at 30 tokens per second, and multiple phone manufacturers confirmed they will adopt it in next-generation flagship phones. The supply of on-device computing power is catching up, which is good news for StepFun, which is betting on devices.

At present, StepFun has not disclosed revenue data, and what can be evaluated is only model reputation, financing progress, and industrial cooperation. According to reports, it has completed shareholding reform and dismantled its red-chip structure, and is preparing for a Hong Kong listing; revenue and profit data will need to wait for disclosure in the prospectus.

Kimi: two questions to be answered

The most recently released model among the five comes from Moonshot AI. Kimi K3.1 Pro has 2.8 trillion total parameters, activates 104 billion per inference, and supports a 1-million context. It was not retrained from scratch; instead, based on K3 released in July, it used a substantially increased amount of compute for post-training. According to the official statement, the most obvious improvements are in coding, 3D creation, financial research, and long-horizon Agent tasks.

The officially announced results are quite impressive: on the Terminal-Bench 4.0 terminal-operation benchmark, K3.1 Pro scored 65.7%, higher than GPT-6 Astra's 58.2% and less than one percentage point behind Claude Opus 5.5's 66.4%. The third-party evaluation firm Artificial Analysis gave it an intelligence index score of 55, 11 points higher than K3, ranking it fourth globally and the highest-ranked open-source model, while its price is only one-fifth that of GPT-6 Astra.

However, the more critical variable in this release is not the benchmark score, but the price.

K3.1 Pro is a post-training upgrade of K3, and its pricing follows the high-end route established by K3. Whether this route can hold depends on a question that has not yet been fully validated: who is continuously paying for the most expensive domestic model. The evidence available so far is that, according to media reports, its ARR rose from $300 million in June to nearly $1 billion by September, and the company's target for the end of the year is $2 billion.

K3.1 Pro's positioning is thus clear: defend the high-end pricing while lowering the entry threshold. Its API price is 13 yuan per million input tokens and 68 yuan per million output tokens, about one-third lower than K3. The high-speed version costs twice as much as the standard version and delivers output two to three times faster.

Artificial Analysis found that K3.1 Pro outputs an average of about 55,000 tokens to complete each test task, 14% more than K3. With the unit price down by one-third and usage up by 14%, the per-task cost is almost unchanged. K3.1 Pro costs about $2.11, versus about $2.00 for K3. For developers, the same money buys stronger capability; for a market expecting "lower prices to drive revenue growth," changes on the billing side will take longer to show.

For Moonshot AI itself, the more direct test lies in the capital market. Moonshot AI's last funding round valued it at about $50 billion, and it plans to list in Hong Kong in the first quarter of 2027. The revenue curve supported by its high-end positioning is the core support for that valuation.

The full weights of K3.1 Pro are promised to be released before November 4. More noteworthy than the weights themselves is a measure the company explained in advance: during the mid-training and reinforcement learning stages, network attack and defense-related data was identified and removed from the training corpus and reinforcement learning environments; training content on vulnerability discovery and exploitation was removed, while general programming tasks were retained. On Moonshot AI's self-developed vulnerability exploitation benchmark, an unfiltered version from the same source solved 90.5% of tasks, while the released version dropped to 81.0%. The official principle is not to release open-source weights that would push cyberattack capabilities to the absolute frontier.

The industry context for this approach is that closed-source vendors can ban APIs and update filter layers when abuse is discovered, whereas once open-source weights are distributed they cannot be recalled, so capability control can only be moved forward into the training stage. According to media observations, proactively removing a category of training data before open-sourcing and accepting a quantifiable capability loss has no precedent among domestic open-source models. The company also acknowledges that as general programming capability improves, safety capability will partially generalize back, and it plans to gradually increase the network capability in open weights, because locally auditable weights are more valuable to defenders.

The next batch of foundation models and a new paradigm

Beyond the five, major companies such as Alibaba, ByteDance, and Tencent also have corresponding moves, though by comparison they have all invested their resources in next-generation foundation models.

Tencent officially announced at its Global Digital Ecosystem Summit on October 29 that the full version of Hunyuan Hy4 is about to launch, adding visual understanding capability for the first time. According to industry media reports, Alibaba's Qwen4 has entered the final sprint to release and opened developer testing, Zhipu's GLM-5.5 has entered the post-training stage, and the next-generation video generation model will still debut in November as planned and announced at the Yunqi Conference. On ByteDance's side, Doubao 2.2 was originally scheduled to launch in August and had previously been confirmed delayed; the Seed team has lengthened the R&D cycle to focus on polishing coding and Agent capabilities.

Another type of change comes from the technical route itself. On October 28, Diffusion Intelligence released the open-source model Clover, which does not follow the old path of autoregressive token-by-token generation but instead uses a diffusion route with parallel decoding, with 35 billion total parameters and 3 billion activated. In the official demo, generation speed reaches up to 10 times that of autoregressive models of the same size. For now it does not constitute a threat, but it shows that beyond the autoregressive route, the technical route for foundation models has not yet converged.

In the same week, overseas developments also provided a reference. Google released Gemini 4 Carbon on October 28, which topped Artificial Analysis's just-revised new intelligence index with 59 points, surpassing Claude Opus 5.5 and GPT-6 Astra. In the old index before the revision, K3.1 Pro ranked fourth globally and first among open-source models with 55 points, 3 points behind the then-leader Claude Opus 5.5. The two rankings use different methodologies, so the scores cannot be directly aligned, but the direction of the gap between open-source and closed-source models narrowing is consistent across both versions.

One can also discern a further trend: the criteria for assessing industry competition are shifting from token consumption to task delivery. DeepSeek's open-source Super-agent architecture is designed with task completion as its goal, with Workers executing in parallel and failed tasks automatically taken over and restarted by a supervisor program, reducing manual intervention during execution.

The business logic behind this trend is that the current token-based billing measures the process of using the model, whereas what customers procure is delivery results. This misalignment in the two metrics means that the current charging method is inconsistent with how customers actually obtain value. Candidate forms of pricing models have already emerged overseas. OpenAI offers some large customers the option of paying per completed task; Salesforce's Fin bills per agent successfully resolved problem; and Sierra charges customers only when AI completes tasks independently without human intervention. The three companies' paths from process-based pricing to outcome-based pricing differ.

Domestic exploration is mainly in hybrid forms. Bairong Yunchuang's outcome-based platform simultaneously offers three models: per-task pricing, per-position pricing, and revenue sharing based on value creation. Researchers at Peking University HSBC Business School summarize this as paying for verifiable incremental business value created by AI agents, and caution that combining process-based pricing such as subscriptions and milestone payments with outcome-based revenue sharing is currently the more realistic mainstream approach. But the difficulty in implementation lies in what constitutes a billable result: there is still no industry-recognized measurement and audit standard, business outcomes arise from multi-party collaboration, value is hard to attribute to a single model call, and result verification relies on after-the-fact data, creating a cycle mismatch with the immediate settlement of tokens.

Some researchers expect that token-based billing will not exit in the short term, and that a full shift to outcome-based payment will take three to five years; the more likely structure is a layered coexistence of the two.

Against this progress, looking again at the moves of leading domestic players—incorporating execution time into core metrics and open-sourcing task-oriented architectures ahead of schedule—can be seen as preparation for the transition to a task-delivery metric. If this trend continues, the pricing basis of the large-model industry will move from single token-based billing toward a layered structure in which tokens and outcomes coexist.

In the next two to three months, two types of milestones will test the above judgment.

On the product side are the next batch of foundation models: the full weights of Kimi K3.1 Pro will be released before November 4, Alibaba's next-generation video model is planned for release in November, Qwen4 has entered the sprint to launch, and the official version of Tencent Hunyuan Hy4 has been officially announced as imminent.

The capital market has already priced this in advance: Zhipu and MiniMax both listed on the Hong Kong Stock Exchange in the same month of January this year, and in February both market capitalizations exceeded HK$300 billion. By the end of October, MiniMax's market capitalization had fallen to less than HK$60 billion, and Zhipu's share price had retreated more than 80% from its June high. The basis for the repricing is MiniMax's interim report, with a gross margin of 17.8% and an adjusted loss of US$293 million. This round of retreat is part of the overall valuation digestion of the AI sector; the fundamentals of the two companies have not stalled—Zhipu's GLM-5.5 has entered post-training, and MiniMax's M3.1 Flash has just cut Agent costs by 70%.

The three unlisted companies will face the same set of tests: Moonshot AI plans to list on the Hong Kong Stock Exchange in the first quarter of 2027, DeepSeek's 100 billion yuan financing has landed and its STAR Market coaching is underway, and StepFun is preparing for a Hong Kong listing. Once prospectuses are filed, each company's revenue structure, gross margin, and losses will be fully disclosed for the first time. Selling capability at a high price, selling scale at a low price, selling to industries, selling to end devices—which of the four routes will work will be answered by the financial reports at that time. (This article was first published on TMTPost APP; author | AGI-Signal; editor | Hao Jingyu)