Summary of Key Points
After much deliberation, the official version of DeepSeek-V4-Pro has been launched, featuring long-duration tasks and enhanced Agent capabilities (supporting up to 1M in context, 384K output, and multiple levels of thinking intensity). However, the API adopts a peak-valley pricing model, which results in increased costs. Compared to its competitor Kimi K3, both models have their strengths and weaknesses: DeepSeek excels in storytelling and rapid modification, while Kimi is more stable in report generation, visual restoration, and physical simulation. Meanwhile, the company behind DeepSeek is undergoing significant changes—expanding its team, securing external funding for the first time, and laying out the infrastructure for Agent applications. The future success of DeepSeek will depend on whether it can maintain its cost-effectiveness and integrate into the daily workflows of ordinary users.
Detailed Analysis
1. New Model: Focused on “Long-Duration Tasks + Intelligent Assistance”
The main selling point of the official DeepSeek-V4-Pro is its ability to handle complex, time-consuming tasks:
- Extensive Memory Capacity: With a context capacity of 1M, it can process large amounts of information at once, similar to handling a thick book or an entire small code library without the need to feed data to the model in multiple rounds.
- Thought-Provoking Output: The model comes with a default thinking mode that can be adjusted to low, high, or max intensity, allowing for more thorough consideration before producing results. Higher settings are particularly useful for complex tasks such as solving math problems or writing complex code.
- Native Integration with Tools: It directly integrates with the Responses API, enabling it to interact with external tools like weather checkers and data calculators, making it ideal for use as an intelligent assistant (Agent).
These improvements are designed to enable the model to handle more substantial work beyond simple conversations.
2. Price Increase: How Does the Peak-Valley Pricing Affect Users?
The DeepSeek API no longer has a fixed price but charges differently during peak and off-peak times:
- Peak Hours: The cost is twice as high between 9 AM and 12 PM and 2 PM (during business hours) compared to off-peak times.
- Price Changes: For example, the cost per thousand tokens has increased from 3 yuan to 4.5 yuan during off-peak hours, and from 6 yuan to 13.5 yuan during peak hours. This means that the cost for output has nearly tripled during peak times.
Why the Price Increase? Long-duration tasks and Agent functions require more computational power, leading to higher costs. How to Mitigate? Developers can schedule their calls during off-peak periods (e.g., running batch tasks at night) to save money. Although DeepSeek’s prices are still lower than those of some top models, its advantage as a “low-cost option” has diminished, and it will need to demonstrate its value through its performance.
3. Head-to-Head Comparison with Kimi K3
A comparison of these two flagship models targeting similar use cases reveals interesting differences:
- Content Creation: DeepSeek is better at storytelling (it continues AI-generated manga more smoothly with less artificial flair), while Kimi excels in report generation (providing tables and data sources for industry analysis).
- Programming Tasks: DeepSeek can make quick adjustments (e.g., modifying a game level in 5 minutes), whereas Kimi offers accurate visual recreations (e.g., creating a webpage based on museum screenshots with consistent structure).
- Visual Quality: Kimi achieved a successful result immediately (creating a jellyfish lake using Three.js with golden jellyfish appearing like a “galaxy”), while DeepSeek experienced multiple failures (blank screens, rendering errors, and jellyfish appearing as gray ellipses).
- Physical Simulation: Kimi’s simulations are logical and realistic (e.g., fireworks exploding without breaking the scene boundaries with proper chain reactions), whereas DeepSeek’s simulations were flawed (fireworks passing through boxes, explosions lacking a clear source of ignition, and lagging).
In summary, DeepSeek is suitable for quick code modifications and storytelling, while Kimi is more reliable for report generation, visual restoration, and physical simulation tasks.
4. DeepSeek’s Transformation: From a Model to an Ecosystem
DeepSeek has evolved from a “lightweight” model, funded by venture capital, to a more comprehensive platform:
- Team Expansion: The company is hiring extensively, doubling the size of its teams and adding positions in areas such as Agent frameworks and data engineering, indicating plans for more advanced products.
- Funding Progress: It has secured external funding for the first time this year (over 50 billion yuan), indicating a need for additional resources to support growth.
- Agent Infrastructure: The launch of the open-source DeepSeek Harness framework allows developers to combine models, tools, and storage to create intelligent assistants (e.g., automating report generation for businesses).
This transition signifies that DeepSeek aims to become the foundation of a larger Agent ecosystem rather than just selling model APIs.
5. Future Challenges: Cost-Effectiveness and User Integration
DeepSeek faces two key challenges:
- Maintaining Cost-Effectiveness: As computing costs continue to rise, if its model capabilities do not significantly outperform others, users may opt for cheaper alternatives. DeepSeek needs to improve execution speed and task completion accuracy to retain users.
- User Adoption: Currently, the DeepSeek Harness framework is designed for developers; ordinary users have limited access to its features. To become a mainstream product, it must offer products tailored for businesses or individuals (e.g., automatically generating weekly reports or handling contracts) to gain real user feedback and turn technical advantages into market benefits.
In conclusion, DeepSeek-V4-Pro is a powerful model, but the price increase and strategic expansion present new challenges. For users, choosing the right tool for their needs (e.g., using DeepSeek for coding during off-peak hours and Kimi for reports) is the most cost-effective approach for now. For DeepSeek, success will depend on its ability to deliver value while managing costs and integrating into everyday user workflows.