Chinese LLM developers are shifting from headline API prices to engineering efficiency, task costs, and business models that can turn rapid adoption into profits.
China’s large language model (LLM) industry has spent more than two years fighting a price war. Now, the competition is shifting toward a more fundamental question: how much does it actually cost to get AI to complete a task?
In May 2024, DeepSeek-V2 cut its input price to 1 yuan ($0.15) per million tokens, prompting ByteDance, Alibaba (9988.HK), and Baidu (9888.HK) to follow suit within a week. Prices continued to fall. In February 2025, Baidu made its Ernie chatbot free, while its Ernie 4.5 model, released the following month, priced its application programming interface (API) at just 1% of the cost of Open AI’s GPT-4.5.
By the second half of 2025, the normalized average price of Chinese LLM token services had fallen to 6.8 yuan, according to the China Academy of Information and Communications Technology. The figure reflects normalized launch prices, however, not what customers actually pay.
A different kind of competition emerged in October. Between Oct. 15 and 24, StepFun, DeepSeek, MiniMax (1011.HK), iFlytek (002230.SZ), and Moonshot AI released new models. None launched a fresh price war. Instead, they focused on reducing the cost of delivering results.
The shift marks the end of competing solely on API prices,. The real cost battle is only beginning.
From price cuts to efficiency
The market began differentiating between models rather than pushing prices uniformly lower. In the first half of 2026, cheaper models and more expensive flagship offerings coexisted. Xiaomi (1810.HK) cut prices on its MiMo-V2.5 series by as much as 99% in May, while several other conversational models followed with cuts exceeding 90% in June.
Yet the normalized average price rose to 13.46 yuan, up 97.9%, as premium models pushed the market higher.
Zhipu (Z.AI) (2513.HK) raised prices three times this year, with CEO Zhang Peng saying the company was returning to normal commercial value. DeepSeek introduced peak and off-peak pricing for its V4 series, raising peak-hour output prices from 6 yuan to 27 yuan.
Moonshot AI’s K3, released in July, pushed output pricing to 100 yuan, making it China’s most expensive domestic model. Its annualized recurring revenue (ARR) reportedly rose from $300 million in June to nearly $1 billion in September.
The October releases confirmed the split. Four of the five companies maintained their previous pricing; only Kimi cut its rates. But the more significant changes were in the cost of completing tasks.
DeepSeek: cutting bills without cutting prices
DeepSeek’s V4.1 Flash-1017, released on Oct. 17, improved efficiency without changing its price. Artificial Analysis found that its September predecessor generated an average of 89,000 tokens per intelligence-index task, 62% more than the previous V4 Pro. Excessive output was eating into its price advantage.
The updated model reduced average task costs from $0.40 to $0.26 at the highest reasoning setting. Running the same benchmark with Moonshot’s K3.1 Pro costs about $2.11, roughly eight times as much.
DeepSeek is also targeting infrastructure and labor costs. On Oct. 20, it open-sourced DeepInfer, an inference framework compatible with Nvidia and Huawei Ascend chips. The company reported throughput of 155 trillion tokens a day on a single cluster.
On Oct. 31, its Harness team released a prototype Super-agent architecture. A central router coordinates parallel workers, while an external supervisor automatically restarts or takes over failed tasks. Early experiments showed half of tasks completed within 12.4 minutes and 95% within 15.7 minutes.
The architecture remains experimental, using synthetic workloads. But its aim is significant: reducing human intervention, which becomes a major expense once token costs have been squeezed.
MiniMax: speed as a competitive advantage
MiniMax’s M3.1 Flash, released on Oct. 19, replaced several full-attention layers with sparse attention. The company says inference speed increased nearly threefold and costs for agent tasks fell 70%. API prices remain 1 yuan per million input tokens and 4 yuan per million output tokens.
The model scored 61.7 in OSWorld V2 computer-operation testing, the highest among domestic models, and tied Claude Opus 5 in BankerToolBench.
In one reported coding comparison, MiniMax Code with M3.1 Flash completed a task in about 55 minutes, compared with 121 minutes for OpenCode paired with DeepSeek V4.1 Flash. Both cost roughly 6.5 yuan.
For MiniMax, efficiency is essential to profitability. The company’s first-half revenue reached $117 million, up 283% year on year, with its open platform and enterprise services accounting for more than 60%. But gross margin was just 17.8%, while adjusted net losses reached $293 million and R&D spending totaled $297 million.
Selling control, not cheaper tokens
iFlytek is pursuing a different strategy with Spark X2.6, combining a cloud model with an open-source 4-billion-parameter on-device model. Its target is to handle 60% of applications locally, 30% through cloud-device collaboration and just 10% through cloud models.
The approach is aimed at education, healthcare and legal services, where data security can matter more than API pricing. iFlytek says the model was trained entirely on domestic computing hardware, with specialized engineering improvements raising long-context training efficiency fourfold and inference efficiency 4.5-fold.
Third-quarter revenue rose 7.99% to 6.56 billion yuan, but R&D spending exceeded 1.7 billion yuan and adjusted losses reached 126 million yuan. The company has established a subsidiary, Token Spark, to separately account for general-model development and commercialization.
StepFun, meanwhile, is betting on devices. Its open-source Step 5 Preview has 600 billion total parameters, activates 27 billion per token and supports a one-million-token context window. The company estimates task costs at one-eighth those of Claude Opus 5 at comparable intelligence.
Its investors include smartphone manufacturers and hardware suppliers, while Tencent Cloud is collaborating with StepFun on automotive AI agents. StepFun has not disclosed revenue and is reportedly preparing for a Hong Kong listing.
Kimi: premium pricing faces a test
Moonshot AI’s K3.1 Pro, released in October, builds on K3 through additional post-training rather than a new pretraining run. Artificial Analysis gave it an intelligence score of 55, ranking it fourth globally and first among open-source models.
Its API costs 13 yuan per million input tokens and 68 yuan per million output tokens, about one-third less than K3. Yet the model generates 14% more tokens per task on average. Artificial Analysis estimates task costs at $2.11, compared with $2 for K3.
The result is stronger performance for roughly the same bill, but not a significant reduction in customer spending.
Moonshot’s premium strategy faces a broader test. Its last funding round valued it at about $50 billion, and it plans to list in Hong Kong in the first quarter of 2027.
The company also plans to release K3.1 Pro’s full weights by Nov. 4, after removing cybersecurity-related training data to limit offensive capabilities. The decision highlights the additional responsibilities associated with releasing open-source models.
From token billing to task delivery
The industry’s next challenge looks likely to be changing how AI is sold. Token billing measures usage, while customers ultimately want completed work.
International providers are already experimenting with outcome-based pricing. OpenAI offers some large customers task-completion pricing; Salesforce’s Fin charges for successfully resolved issues, while Sierra bills when AI completes tasks independently.
Chinese companies are exploring hybrid approaches, combining task-based fees, subscriptions, and revenue-sharing. But there is no widely accepted standard for measuring or auditing results, making a full shift away from token pricing difficult.
Researchers expect token billing to remain dominant in the short term, with a broader transition taking three to five years.
The next few months will test the economics of the industry. DeepSeek, StepFun and Moonshot AI are preparing for potential listings, while MiniMax and Zhipu’s market valuations have fallen sharply from their peaks.
Ultimately, the financial statements will reveal which business models can work: selling premium capabilities, low-cost scale, industry-specific solutions or AI embedded in devices.
The price war may be over. The cost battle will determine who survives.
Source:
TMT Post
