Intel CEO Lip-Bu Tan said the performance of Chinese AI models such as DeepSeek, Kimi and GLM is "good enough" at a panel discussion hosted by DBS in Singapore on October 12, 2026. According to Reuters, he noted that advanced Chinese AI models charge less per token and are being adopted rapidly. His view is that what matters is not only competing on peak model performance, but how cheaply users can access the capability each use case requires. However, not all Chinese AI models are equally cheap. Comparing each company's official pricing shows large differences between models, and actual costs also vary depending on the time of day and on whether input data hits the cache.
Intel CEO Calls Chinese AI Models "Good Enough"
According to Reuters (distributed via Euronext), Tan said adoption of Chinese AI models has been "very fast" and that the market will decide which model suits which use case. On open source, he said, "I strongly believe in open source. It is the best way to improve efficiency," stressing the importance of an open technology foundation for developing and using AI.
That said, this is a market assessment from a semiconductor executive, not proof that Chinese AI models match competitors' products in every use case. The Reuters article does not provide specific adoption-rate trends or benchmark results comparing the models under identical conditions. Tan also did not specify which generation of DeepSeek, Kimi or GLM models he meant, so caution is needed in directly linking his remarks to each company's current products and pricing.
Chinese AI Model Pricing Shows Large Gaps
Comparing the official API pricing for DeepSeek, Kimi and Z.AI, as confirmed on October 15, 2026, shows that prices differ greatly even among Chinese AI models. The table below shows prices per million tokens in US dollars. Tokens are the units AI models use to process text, and both the text you input and the text the model generates are billed. Input prices use the rate that applies when previously processed content cannot be reused from the cache.
| Model / time period | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek V4.1 Flash, off-peak | $0.15 | $0.60 |
| DeepSeek V4.1 Flash, peak | $0.30 | $1.20 |
| Kimi K3 | $3.00 | $15.00 |
| GLM-5.3 | $1.40 | $4.40 |
| GLM-5.3-Flash | $0.15 | $0.50 |
This table compares API prices only. It does not account for model performance or for differences in the amount of processing needed to complete the same task. It also excludes extra charges, such as cache writes or the use of external tools.
For example, processing 1 million uncached input tokens and 1 million output tokens on DeepSeek V4.1 Flash costs a total of $0.75 during off-peak hours. During peak hours, the same workload costs $1.50, so the time of use alone doubles the price. This is an estimate based on official list prices, not a measured cost of completing a specific task.
Kimi K3's output price of $15 per million tokens also differs sharply from DeepSeek V4.1 Flash and GLM-5.3-Flash. Within the GLM series, too, the standard GLM-5.3 and the Flash version are priced differently, so being Chinese-made does not by itself mean a model is cheap. When estimating actual deployment costs, you need to specify the exact model and check both input and output prices.
DeepSeek's Price Doubles at Peak Hours, and Caching Also Makes a Difference
DeepSeek's API pricing sets different rates for peak and off-peak hours. Peak hours are 1:00–4:00 and 6:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. In Japan Standard Time, that corresponds to 10:00 a.m.–1:00 p.m. and 3:00–7:00 p.m. Off-peak rates apply at all other times, on weekends and on Chinese public holidays. In other words, the cost of the same model can differ significantly between real-time use during the Japanese daytime and batch processing at night.
Input caching is another key factor affecting costs. For DeepSeek V4.1 Flash, the off-peak input price per million tokens falls from $0.15 on a cache miss to $0.003 on a cache hit. For workloads that repeatedly use shared instructions or reference materials, raising the cache hit rate can cut input costs substantially. However, this discount applies only to the portion of input reused from the cache, and output costs do not fall by the same proportion.
Kimi K3 also offers caching, but its pricing structure differs from DeepSeek's. Under Kimi's official pricing, cache writes cost $3 per million tokens for a five-minute validity period and $6 for one hour. Cached input hits cost $0.30 per million tokens, and a hit extends the validity period without incurring an additional write charge. This means the cost of creating a cache and the savings from reusing it need to be considered separately.
When external tools are used, token prices alone also don't give the total cost. For example, Z.AI charges $0.01 per use of its web search tool. For workloads that combine an AI model with search or other processing, the number of external tool calls affects costs in addition to the token count. Simply choosing the cheapest model on the API price list does not necessarily minimize actual operating costs.
The Design Behind DeepSeek's Low Prices, and the Challenges of Self-Hosting
In its announcement of DeepSeek V4.1 Flash on September 10, 2026, DeepSeek described a new architecture designed to improve processing efficiency. The model uses a Mixture of Experts (MoE) approach, which computes only the necessary parts of the model. Of 552 billion total parameters, 8 billion are active when processing input and 16 billion when generating output. This asymmetric design, which uses different computing resources for input and output, aims to balance processing capacity with efficiency.
DeepSeek also says that, compared with the previous generation, it reduced the HBM capacity required for the KV cache to one-quarter and the SSD capacity to one-eighth. The KV cache stores computation results generated while an AI model processes text so they can be reused in later processing. Reducing the memory and storage required could lower equipment costs when handling large volumes of requests. However, these reduction rates are figures published by DeepSeek itself and have not been confirmed by independent verification.
On the other hand, being able to use an API cheaply is a separate matter from being able to run the same model on your own equipment at low cost. In its V4.1 Flash announcement, DeepSeek invites companies considering large-scale deployments using 2,000 GPUs and a storage cluster to contact it. This is not a minimum configuration for running the model, but it suggests that large-scale self-hosting requires substantial equipment. You cannot estimate the cost of running the same model on your own infrastructure based solely on API token prices.
For Intel, too, wider adoption of publicly available models could create opportunities to make use of its computing infrastructure. In an official technical explainer, the company describes how to run distilled DeepSeek R1 models on Xeon processors and Gaudi accelerators. Distillation is a technique that uses the outputs of a large model to train another model to learn its capabilities, and the models discussed there are different from DeepSeek V4.1 Flash itself. Tan's support for open source is consistent with Intel's policy of supporting the use of public AI models on its own hardware, but whether it will translate directly into higher sales requires separate verification.
What Matters Is Cost Per Task, Not Price Per Token
When choosing an AI model for business use, it is important to evaluate not just the price per token but also the amount of processing and effort needed to complete a single task. Even a cheap model can end up costing more in input and output tokens if its answers are of insufficient quality and require repeated corrections or regeneration. Conversely, for uses where the required quality can be maintained and the task completed in one pass, a low per-token price is more likely to translate directly into ongoing cost savings.
For example, classifying inquiries and fixing complex software differ greatly in the accuracy required and the impact of failure. Also, even with the same input text, models split tokens differently and generate answers of different lengths, so the ratio of unit prices on a price list does not necessarily equal the ratio of cost per task. To compare models in real work, first define the acceptable quality and response time, then evaluate them including the number of reruns and the burden of human review and correction.
Tan's idea that the market will choose the model suited to each use case will be decided by cost-effectiveness in actual work. The highest-performing model is not necessarily optimal for every task. If a model can meet the required quality and response time while keeping the total cost of completing each task low, a low-priced model with "good enough" performance has clear value for adoption. Going forward, in addition to benchmark performance and per-token prices, how efficiently a model can complete actual work will be an important factor in choosing an AI model.