DeepSeek’s Pricing Strategy: Why Liang Wenfeng Prioritizes Efficiency Over Maximum Profit

Introduction

DeepSeek founder Liang Wenfeng has offered one of the clearest explanations yet of how the company thinks about pricing, cost efficiency, open models, and commercialization.

In remarks attributed to a four-hour investor meeting, Liang said DeepSeek does not set prices to maximize revenue. The company instead aims to earn what he described as a reasonable profit, while keeping its models affordable enough for widespread use.

The most visible example was DeepSeek-V4-Pro. The model was initially priced at a relatively high level because the company expected demand to exceed its available high-end computing capacity. DeepSeek later reduced the price to one quarter of the original level and made the 75% reduction permanent.

Liang’s argument goes beyond a conventional price promotion. He treats low cost as the result of architecture and engineering efficiency—and as a requirement for frontier-model research under limited computing resources.

The meeting comments circulating online were compiled from limited material and are not an official verbatim transcript published by DeepSeek. Some wording may therefore differ from what was said in the room. Current model names, features, and API prices in this article were checked against DeepSeek’s official documentation.

DeepSeek Says It Wants Reasonable Profit, Not Maximum Revenue

The central idea in Liang’s pricing comments is restraint.

Most technology companies are expected to increase revenue, expand market share, build sales organizations, and capture as much value as possible from a successful product. DeepSeek’s stated objective is narrower.

Liang reportedly said:

“We only earn a reasonable profit.”

The statement does not mean that DeepSeek rejects commercial activity. The company charges for API usage, operates consumer products, serves developers, and has continued to expand model access.

The distinction is between commercial operation and commercial maximization.

DeepSeek appears willing to earn enough revenue to operate services and support development, but it does not currently want pricing decisions to be dominated by the goal of extracting the highest possible margin from each user.

That philosophy reflects the company’s larger focus on artificial general intelligence. Liang has described products and services as steps along the path to more capable systems, rather than the final destination of the company.

DeepSeek-V4-Pro Was Cut to One Quarter of Its Original Price

DeepSeek-V4-Pro provides a concrete example of this approach.

When the model launched, its API price was set considerably higher than the lighter V4-Flash model. DeepSeek cited constraints in high-end computing capacity and expected the price to fall when more infrastructure became available.

In May 2026, the company made a 75% price reduction permanent.

Reported yuan-denominated prices fell from a range of approximately 0.1 to 24 yuan per million tokens to 0.025 to 6 yuan per million tokens, depending on whether the tokens were cached input, uncached input, or output.

Liang reportedly explained that V4-Pro had initially been priced high because DeepSeek was concerned about excessive demand. Once it was possible to reduce the price, the company cut it directly to one quarter.

According to the meeting account, employees celebrated the reduction because the purpose of building the model was to let more people use it fully.

This framing is unusual. A price cut is commonly described in terms of customer acquisition, competitive positioning, or market-share growth. Liang instead presented it as a better alignment between the model’s purpose and its accessibility.

Current DeepSeek API Pricing

DeepSeek’s official API documentation currently lists two main V4 endpoints:

Both models support thinking and non-thinking modes, JSON output, tool calls, chat-prefix completion, and a maximum context length of one million tokens, according to the current documentation.

The listed concurrency limits are higher for V4-Flash than V4-Pro:

• V4-Flash: 2,500
• V4-Pro: 500

DeepSeek also states that prices may change and recommends checking the official pricing page before estimating production costs.

The older deepseek-chat and deepseek-reasoner model names are scheduled for deprecation on July 24,
2026. They currently map to the non-thinking and thinking modes of deepseek-v4-flash for compatibility.

Why DeepSeek Initially Charged More

A high launch price does not always mean a company believes the product’s value has permanently increased.

It can also be a capacity-management mechanism.

When demand is greater than the available supply of inference compute, a provider has several choices:

1. Create a waiting list.
2. Apply strict rate limits.
3. Reserve access for selected customers.
4. Reduce the model’s available features.
5. Charge a higher price to suppress demand.
6. Expand infrastructure and lower the price later.

DeepSeek appears to have used a combination of pricing and capacity limits.

V4-Pro required scarcer high-end computing resources, so a higher initial price reduced the chance that demand would overwhelm the service. When more capacity and efficiency became available, the company could offer the model more broadly.

This is different from setting a high price because customers are willing to pay it.

The first approach treats price as a temporary resource-allocation tool. The second treats price as a way to capture the maximum perceived value.

Liang’s remarks suggest that DeepSeek prefers to lower the price when the capacity constraint is removed.

Low Cost Is an Architectural Result, Not Only a Commercial Tactic

Liang reportedly emphasized that DeepSeek’s low costs are not simply a business strategy.

They are the result of model architecture, systems engineering, and hardware-aware optimization.

DeepSeek has repeatedly focused on techniques intended to improve the amount of useful model capability obtained from a limited computing budget.

Examples across its research include:

• Mixture-of-Experts architectures that activate only part of the model for each token.
• Multi-head Latent Attention to reduce key-value cache requirements.
• Low-precision training.
• Communication-efficient training systems.
• Hardware-aware co-design.
• Context caching.
• Sparse computation.
• Optimized inference and serving.

DeepSeek-V2, for example, used 236 billion total parameters while activating 21 billion per token. Its research paper reported lower training costs, a substantially smaller KV cache, and higher generation throughput compared with the earlier DeepSeek 67B system.

DeepSeek-V3 expanded the approach to 671 billion total parameters and 37 billion active parameters per token. The company reported training it on 14.8 trillion tokens while keeping the API price low relative to many frontier-model services.

DeepSeek-V4 extends the strategy with a one-million-token context window and separate Flash and Pro models.

These systems do not make computing free. Large models still require expensive chips, networking, electricity, engineering, storage, and data-center operations.

The point is that architecture can change how much capability is produced from each unit of compute.

Cost Efficiency Helps DeepSeek Train Larger Models

Liang’s deeper argument concerns research capacity.

A company with effectively unlimited computing resources can improve a model by adding more chips, larger clusters, and more training runs.

DeepSeek does not operate under the same resource conditions as the largest U.S. technology companies. Export restrictions, chip availability, capital requirements, and infrastructure constraints all influence what it can train.

In that environment, lowering the cost of each unit of computation has two effects.

It Reduces the Price of Serving Users

More efficient inference allows DeepSeek to process more requests with the same infrastructure. That can lower API prices and expand access.

It Expands the Model That Can Be Trained

If the training system uses compute more efficiently, the company can train a larger or more capable model within a fixed hardware budget.

Liang reportedly summarized the difference this way: large companies can solve problems by adding resources, while DeepSeek prioritizes cost efficiency.

This is why low pricing and frontier research are connected in DeepSeek’s strategy. The same engineering work that reduces inference cost may also help the company move further with limited training resources.

Cost Comes Before Time and User Experience

In the broader meeting account, Liang reportedly identified three important factors in frontier-model competition:

1. Cost.
2. Time.
3. User experience.

Cost was placed first.

When two companies offer models of similar quality, the provider that can serve the model more cheaply has a structural advantage. It may charge less, support more users, run more internal experiments, or reinvest the savings into the next training cycle.

Time comes next because a lead of several months can matter in a rapidly changing market. Models, hardware, agent frameworks, and customer expectations move quickly.

User experience can create loyalty, but Liang reportedly did not describe it as the most fundamental technical moat.

This ordering helps explain why DeepSeek invests heavily in architecture and infrastructure while keeping its sales and product organization relatively restrained.

DeepSeek Runs Its API Business With a Small Team

Liang reportedly said he did not find the API business itself especially attractive.

That comment may sound surprising because API revenue is a major business model for frontier-model providers.

DeepSeek’s position is easier to understand when viewed through its operating model.

According to the meeting account:

• A small team can maintain the service.
• The company has very limited customer support.
• It does not need a large sales organization.
• Users discover and integrate the models on their own.
• Product demand is driven by model capability and price.

This is a product-led distribution model.

Instead of hiring a large enterprise sales force to negotiate contracts, DeepSeek publishes models, provides compatible APIs, offers documentation, and lets developers begin using the service directly.

The approach can keep operating costs low, but it also has trade-offs.

Large organizations often require:

• Procurement support.
• Security documentation.
• Service-level agreements.
• Account management.
• Migration assistance.
• Compliance reviews.
• Dedicated capacity.
• Incident support.
• Integration consulting.

A minimal commercial organization may work well for self-service developers and technically capable teams. It may be less suitable for enterprises that expect extensive support and contractual guarantees.

“We Are Commercializing, but Commercialization Is Not the Goal”

Liang reportedly said that DeepSeek has always been commercializing, but does not treat commercialization as its ultimate purpose.

This is not the same as saying the company has no business model.

DeepSeek currently has several mechanisms that can support revenue or strategic value:

• Paid API access.
• Consumer applications.
• Developer adoption.
• Enterprise usage.
• Model ecosystem growth.
• Open-weight distribution.
• Technology licensing or partnerships.
• Infrastructure relationships.
• Talent attraction.
• Potential future products.

The statement is better understood as a priority hierarchy.

DeepSeek’s first objective is to advance model capability and pursue AGI. Commercial activity supports that effort, but the company does not want short-term revenue goals to redirect research toward every popular product category.

Liang reportedly said the point at which DeepSeek fully shifts toward commercialization remains far away.

That position may be easier to maintain while the company has strong funding, patient investors, or other sources of financial support. As compute needs grow, however, even research-centered laboratories must address the cost of infrastructure, employee retention, and long-term operations.

A Low Price Can Be a Competitive Weapon Even When Profit Is Not Maximized

DeepSeek may not describe its pricing as a market-share campaign, but low prices still affect competitors.

A provider selling a model with competitive quality at a fraction of the prevailing API cost forces the market to respond.

Competitors may need to:

• Lower prices.
• Introduce smaller models.
• Improve caching.
• Offer batch discounts.
• Optimize inference.
• Create cheaper routing tiers.
• Bundle models with cloud services.
• Justify premium prices through tools, reliability, or support.

The result is a broader reduction in the cost of deploying AI.

This benefits developers, startups, researchers, and companies that previously could not afford high-volume model usage.

It can also create pressure on model providers whose economics depend on maintaining high gross margins.

DeepSeek’s price strategy therefore has commercial consequences even when the company says profit maximization is not its objective.

Open Models Are Part of the Same Strategy

Liang’s remarks about pricing are closely connected to DeepSeek’s support for open model releases.

He reportedly described open release as a form of giving up part of the company’s potential benefit.

Publishing model weights allows:

• Researchers to study the architecture.
• Developers to run models privately.
• Companies to fine-tune or adapt them.
• Infrastructure providers to optimize deployment.
• Competitors to learn from the work.
• The broader ecosystem to build without using DeepSeek’s hosted API.

From a narrow commercial perspective, this can reduce API lock-in.

From DeepSeek’s perspective, open models also create:

• Technical influence.
• Developer adoption.
• Research feedback.
• Recruitment value.
• Organizational pride.
• Faster ecosystem innovation.
• A stronger position in global AI discussions.

Liang reportedly said DeepSeek would not publish a deliberately weakened model while keeping a substantially stronger version private.

That claim should be evaluated release by release, but it reflects the company’s public positioning: open access is not only a marketing sample; it is intended to be a meaningful part of the research strategy.

What “Reasonable Profit” Does Not Mean

The phrase “reasonable profit” is attractive, but it is not a precise financial metric.

DeepSeek has not publicly defined:

• A target gross margin.
• A target operating margin.
• A required return on infrastructure.
• How research costs are allocated to API pricing.
• Whether consumer products are subsidized.
• How pricing changes with chip supply.
• How much capacity is reserved for internal use.
• Whether future enterprise services will use different margins.

The phrase should therefore be understood as a statement of intent, not a published pricing formula.

It also does not guarantee that prices will never increase.

DeepSeek’s official documentation states that product prices may change. Capacity constraints, hardware costs, demand, model size, and service requirements can all affect future pricing.

A company can reject maximum profit while still increasing prices when the cost of providing a service rises.

The Difference Between Cheap Tokens and Low Total Cost

API price per million tokens is easy to compare, but it does not capture the full cost of using a model.

A production application should also measure:

• How many tokens the model uses to complete a task.
• Whether it requires retries.
• The quality of tool calls.
• Latency.
• Cache hit rates.
• Context length.
• Output length.
• Error rates.
• Human-review time.
• Reliability under load.
• The cost of switching models.
• The value of support and service guarantees.

A model with a lower token price may cost more overall if it produces long answers, fails frequently, or requires extensive correction.

A more expensive model may be economical if it solves the task in one attempt.

DeepSeek’s cost advantage should therefore be tested at the level of a completed workflow rather than only at the published token rate.

What Developers Should Check Before Choosing DeepSeek API

Model Fit

Compare V4-Flash and V4-Pro on the actual task. Flash may be sufficient for high-volume workflows, while Pro may be more appropriate for difficult coding, reasoning, or agent tasks.

Cache Behavior

Cached input is dramatically cheaper than uncached input. Applications with repeated system prompts, shared documents, or stable context may benefit substantially from caching.

Output Length

Output is more expensive than input on both models. Unnecessarily long reasoning traces or verbose responses can dominate the bill.

Concurrency

The documented concurrency limits differ between Flash and Pro. High-volume products should test capacity requirements early.

Model Names

Developers still using deepseek-chat or deepseek-reasoner should migrate to the V4 model identifiers because the legacy names are scheduled for deprecation.

Provider Risk

DeepSeek reserves the right to change prices. Production systems should avoid hard-coding economic assumptions that depend on one permanent price.

Support Requirements

Self-service API access may be enough for an individual developer. Enterprises should confirm security, compliance, service, and support expectations before standardizing on a provider.

Strategic Advantages of DeepSeek’s Approach

DeepSeek’s pricing model creates several potential advantages.

Wider Adoption

Low prices reduce the cost of experimentation and make the models accessible to smaller teams.

More Usage Data and Feedback

A larger user base produces more information about failures, popular workflows, infrastructure bottlenecks, and developer needs.

Pressure Toward Efficient Engineering

When the company cannot rely on premium margins, it has a strong incentive to improve architecture and serving efficiency.

Stronger Open Ecosystem

Affordable hosted APIs and downloadable weights reinforce each other. Developers can begin with the API and later move to private deployment.

Talent and Research Reputation

Engineers may be attracted to a laboratory that prioritizes frontier work, open releases, and long-term technical goals.

Risks and Trade-Offs

The same strategy also creates challenges.

Infrastructure Demand

Very low prices can attract demand faster than capacity grows.

Limited Enterprise Support

A small sales and support team may slow adoption in highly regulated or procurement-heavy organizations.

Sustainability

Frontier training costs continue to rise. Reasonable profit must still cover service operation and contribute to future research.

Price Expectations

Once developers become accustomed to very low prices, future increases may be difficult.

Competitive Response

Large cloud and model providers can subsidize low-cost tiers, bundle models with infrastructure, or use scale to reduce prices.

Commercial Focus

Avoiding non-core products can preserve research focus, but it can also leave valuable markets to competitors.

常见问题

What did Liang Wenfeng say about DeepSeek’s pricing strategy?

Liang reportedly said DeepSeek aims to earn a reasonable profit rather than maximize revenue. He connected low pricing to the company’s goal of making capable models widely usable and improving cost efficiency.

Did DeepSeek reduce a model’s price by 75%?

Yes. DeepSeek made a 75% reduction for V4-Pro permanent, leaving prices at one quarter of their original level. The exact cost depends on cached input, uncached input, and output usage.

Why was DeepSeek-V4-Pro expensive at launch?

DeepSeek said high-end computing capacity was constrained. The higher initial price helped control demand until more capacity and efficiency became available.

What are the current DeepSeek-V4 API prices?

DeepSeek’s official documentation currently lists V4-Flash at $0.0028 per million cached input tokens, $0.14 per million uncached input tokens, and $0.28 per million output tokens. V4-Pro is listed at $0.003625, $0.435, and $0.87 respectively.

Is DeepSeek against commercialization?

No. Liang reportedly said the company has always pursued commercial deployment, but commercialization is not its ultimate objective. The stated priority remains model research and the long-term pursuit of AGI.

How can DeepSeek offer low API prices?

DeepSeek attributes much of its cost position to model architecture, sparse computation, caching, low-precision training, infrastructure optimization, and a lightweight operating model. Hardware availability and capacity also influence pricing.

Does DeepSeek provide enterprise sales and support?

DeepSeek offers documentation, APIs, usage limits, and capacity-expansion requests, but Liang described the company as operating with a small service team and limited traditional sales or customer-support infrastructure. Organizations should verify whether the available support matches their requirements.

Will DeepSeek API prices remain unchanged?

Not necessarily. DeepSeek’s official pricing page says prices may vary and that the company reserves the right to adjust them. Developers should check the current documentation when estimating costs.

相关工具

• DeepSeek Chat: DeepSeek’s official web interface for interacting with its current models.
• DeepSeek API Platform: The official platform for API keys, usage, billing, and model access.
• DeepSeek API Documentation: Official integration guides, model information, pricing, limits, and change logs.
• DeepSeek on GitHub: Official repositories for model research, inference components, training systems, and open-source projects.
• DeepSeek on Hugging Face: Official open-weight model releases, model cards, and downloadable checkpoints.
• DeepSeek-V4 Collection: Official V4 model collection containing Flash, Pro, base, and deployment variants.

Related Links

• DeepSeek Models and Pricing: The current official API prices, context limits, features, and model names.
• DeepSeek V4 Preview Release: Official information about V4-Pro, V4-Flash, context length, and API access.
• DeepSeek API Change Log: Official updates to models, compatibility names, features, and availability.
• DeepSeek-V4 Hugging Face Collection: Official model weights and associated releases.
• DeepSeek-V3 Release: Official architecture and historical API pricing information for V3.
• DeepSeek-V2 Research Paper: Technical details on DeepSeekMoE, Multi-head Latent Attention, training efficiency, and inference improvements.
• Public Compilation of Liang Wenfeng’s Investor-Meeting Remarks: A compilation attributed to elsewhere that notes some wording may differ from the original remarks.

Summary

Liang Wenfeng’s reported pricing comments show that DeepSeek views cost efficiency as a technical and strategic objective, not only a way to attract customers.

The company initially priced V4-Pro high to manage demand under limited high-end computing capacity, then reduced the price by 75% when it could serve the model more broadly. Current official API prices remain low for both V4-Flash and V4-Pro.

DeepSeek is commercializing its technology, but Liang does not describe maximum revenue as the company’s ultimate goal. Its model combines self-service distribution, a small operating team, open-weight releases, architectural efficiency, and a long-term focus on AGI research.

DeepSeek’s central pricing bet is that the company with the most efficient path to capable models can expand access, compete globally, and fund future progress without extracting the maximum possible price from every token.