Summary of Key Points
Recent commercial initiatives by OpenAI and DouBao highlight a clear divergence in the charging models for large-scale language models both domestically and internationally: overseas companies (such as OpenAI) are accelerating the adoption of token-based pricing at the enterprise level (where each token represents the smallest unit of AI processing, similar to pay-as-you-go model), while maintaining a subscription-based model for individual users with certain limitations; in contrast, domestic companies (like DouBao) primarily rely on a monthly subscription fee without strict limits on the number of usage tokens. This difference is influenced by factors such as cost pressures, user habits, and market competition. The widespread adoption of token-based pricing is driving the industry's valuation paradigm from focusing on the number of users to emphasizing actual usage volume. In the future, China may move towards a hybrid model that combines both subscription and token-based billing.
I. Charging Models at Home and Abroad: One Based on Usage, One on Monthly Subscription
- Overseas Model: Individual subscriptions serve as a foundation, with enterprises transitioning to token-based pricing.
ChatGPT offers free, Plus, Pro, and Business tiers for individual users. The free version limits users to 10 GPT-5.5 messages every 5 hours; Plus and Pro users are limited to 160 messages every 3 hours, after which they switch to the mini version. For enterprises (e.g., Codex), fees are calculated based on the actual number of tokens used for input, output, and caching. Anthropic takes a more aggressive approach, charging $20 per user per month plus additional fees for token usage, with no free quota.
- Domestic Model: Subscription-based pricing is the standard, with no clear limits on usage volume.
DouBao's professional versions come in three tiers: Standard ($68/month), Enhanced ($200/month), and Professional ($500/month), all of which have no limit on the number of messages. Kimi also uses a subscription model; DeepSeek, on the other hand, does not offer a consumer subscription and focuses solely on providing developer APIs.
- Commonality: Both models target heavy users. OpenAI Codex has 40% of its new users coming from non-developer backgrounds (analysts, operators, etc.), indicating that the tool is becoming more widely used in general office settings. Heavy users can benefit from price discounts through subscriptions (for example, paying $200 per month to save on API costs that would otherwise amount to over $2,000).
II. Overseas Companies Rushing to Adopt Token-Based Pricing Due to Cost Pressures
The core logic behind token-based pricing is to precisely reflect the cost of computing power:
- Subscription models have a ceiling: If users' monthly token consumption exceeds the cost of the subscription, companies may incur losses (e.g., when processing long texts or using tools extensively).
- Enterprises represent a gold mine: They consume large amounts of tokens and have strong payment capabilities, making token-based pricing a source of additional revenue. For example, OpenAI generates 21% of its revenue from enterprise customers, and Anthropic's annual recurring revenue reached $47 billion (up from $9 billion last year), primarily from corporate contracts.
- The anxiety surrounding computing power costs is a driving force: AI operations require numerous GPUs, which are very expensive. Token-based pricing directly links costs to revenue, helping companies manage their computational expenses more effectively.
III. Domestic Companies Adhere to Subscription-Based Pricing for Fear of Losing Users and Intense Competition
Domestic companies are hesitant to switch to token-based pricing for practical reasons:
- User Perception: Paying per use (with tokens) makes users feel like they are spending money for each individual service, which may deter them from using the product. A subscription model is more akin to a monthly subscription package and is psychologically more acceptable to users.
- User Base Is Crucial: DouBao has 345 million active monthly users, making it the largest AI application in China. Companies are reluctant to disrupt this user base with aggressive pricing strategies.
- Fierce Competition: Many AI apps in China are still free, so a subscription model is considered a relatively mild way of generating revenue. If companies switch to token-based pricing, users might switch to free alternatives.
- Ecosystem Development: DouBao is part of larger ecosystems such as ByteDance's Lark (for enterprise collaboration) and Koozi (for AI programming). These companies aim to retain users through subscriptions before gradually transitioning them to paid ecosystem services.
IV. The Major Changes Token-Based Pricing Will Bring to the Industry
Token-based pricing is not just a new method of charging but also a reconstructor of the industry's valuation and supply chain:
- Changing Valuation Models: Traditional internet companies focused on the number of users multiplied by ARPU (Average Revenue Per User); now, large-scale language models are valued based on effective token consumption multiplied by the value of each token, as tokens directly reflect computing power costs and service value.
- Opportunities in Various Industry Segments:
1. Computing Power: The more tokens are used, the more computing power is required, benefiting data centers and cloud computing rental services.
2. Telecommunications: Token transactions require data transmission, boosting the demand for communication networks and related services.
3. Applications: A hybrid model (subscription + token) makes revenue generation more flexible, allowing companies to optimize their revenue structures by combining subscription-based and token-based pricing strategies.
V. Price Divergence: More Valuable Tokens in Real-World Applications
The divergence in pricing patterns between domestic and overseas markets is driven by the value of different use cases:
- Overseas: Newer models are more expensive (e.g., GPT-5.5 is twice as costly as its predecessor). Anthropic has even implemented hidden price increases by using new tokenizers, increasing the cost of the same amount of text processing by 40%.
- Domestic Market:
- General-purpose API prices have decreased (e.g., DeepSeek and Xiaomi have reduced their API prices by over 90% to attract more users).
- Productivity-related services have seen price increases (e.g., Zhispu GLM-5 and Tencent CodeBuddy, with price hikes of over 100%) because these tools directly contribute to users' productivity and are therefore more profitable.
Future Trends: A Hybrid Model May Become the Norm
The combination of token-based pricing overseas and subscription-based models in China is not about one replacing the other but represents different choices at different stages of development. Overseas companies have mature ecosystems that can afford usage-based pricing, while domestic companies with large consumer bases prefer to retain users through subscriptions before gradually transitioning to token-based billing for enterprises. In the end, a hybrid model (subscription for consumers + token-based pricing for enterprises) is likely to emerge, balancing user habits and corporate cost considerations.
The competition in AI commercialization boils down to determining how much users are willing to pay for AI services, and tokens play a key role in measuring this. Companies that successfully balance user experience and costs will emerge as winners in this market.