Don't worry about Claude's supply cutoff. Doubao's programming model is here. It can create a replica of "Minecraft" in 5 minutes at a cost of only 0.2 yuan.
The first Doubao programming model is here!
According to a report by Zhidongxi on November 11th, today, Volcengine, the cloud and AI service platform under ByteDance, released the first programming model in the Doubao large model family - Doubao-Seed-Code. This is a programming model specifically optimized for Agentic Coding tasks, and it has achieved a breakthrough in cost - effectiveness.
In terms of performance, in several mainstream programming evaluation sets in the industry, Doubao-Seed-Code scored higher than domestic models such as DeepSeek-V3.1, Kimi-K2, and GLM-4.6. Its overall performance is only second to the current top model in the field of AI programming - Claude Sonnet 4.5. In addition, Doubao-Seed-Code has a native 256K context, which is higher than the 200K context of Claude Sonnet 4.5.
Beyond the rankings, Doubao-Seed-Code also focuses on implementation in real programming scenarios. Thanks to its optimization for mainstream development tools, users of Claude Code, Trae, or veCLI can easily get started and obtain stable output results.
Meanwhile, Doubao-Seed-Code is the first domestic programming model that supports visual understanding ability. It can generate code by referring to UI design drafts, screenshots, or hand - drawn sketches, or conduct visual comparisons of the generated pages, and independently complete style repairs and Bug fixes, greatly improving the efficiency of front - end development.
It is worth mentioning that today, the Chinese version of Trae, an AI - native IDE product under ByteDance, has integrated the Doubao-Seed-Code model. The combination of Trae + Doubao-Seed-Code topped the SOTA in the authoritative programming benchmark test SWE - Bench - Verified, achieving an ecological closed - loop of model + tool.
Considering the price, Doubao-Seed-Code becomes even more competitive. This model adopts a tiered pricing model. In the 0 - 32K input range where users use the most, the input price of Doubao-Seed-Code is 1.20 yuan per million Tokens, and the output price is 8.00 yuan per million Tokens. After using the full - volume transparent cache, the usage cost of the model can be further reduced by 80%, and the comprehensive usage cost is reduced by 62.7%.
During our actual test, we asked Doubao-Seed-Code to reproduce the classic game "Minecraft", and the cost was less than 0.2 yuan. The painting style and game experience are similar to the original version, and it can be played directly. With the same amount of tokens (in the 0 - 32K range), the cost of Claude Sonnet 4.5 to complete the task has exceeded 3 yuan.
Recently, Zhidongxi had an in - depth early experience of the effectiveness of Doubao-Seed-Code in real programming scenarios. It can not only independently plan development plans and quickly build front - end web pages, but also make in - depth modifications to the database. When encountering errors, it can actively repair, add comments, and optimize the structure. At this moment, Doubao-Seed-Code is no longer just a "code - writing machine", but a development partner who can think side by side with people and create together.
01. Seamlessly integrate with Claude Code, and reproduce "Douyin" with just one screenshot
Tool compatibility has always been one of the main factors affecting the adoption of AI programming models. This time, Doubao-Seed-Code has made great efforts in tool adaptation. Doubao-Seed-Code chooses to be natively compatible with the Anthropic API and can directly integrate with Claude Code without any conversion. This means that a large number of developers familiar with Claude Code can easily switch the API to Doubao-Seed-Code with almost no learning cost.
The Volcengine Ark platform also provides a detailed API call guide for Doubao-Seed-Code. Even novice users can experience the new model by following the guide step by step.
Before conducting an in - depth actual test of the development ability of Doubao-Seed-Code, let's start with a few "appetizers".
The bouncing ball has almost become a must - test question for large models. The bouncing ball created by Doubao-Seed-Code not only conforms to the laws of physics but also runs very smoothly. At the same time, the model independently decided to add a new function. As long as you click on the ball, you can change its force situation, making the ball's bounce no longer just an infinite loop.
We also tested the model's ability to develop based on web page screenshots. After uploading the screenshot, Doubao-Seed-Code can analyze the page layout, visual features, etc., and then gradually build the core components of the web page. Before sending the result to the user, the model will also conduct a functional test and then deliver the finished product.
In daily application scenarios, using Doubao-Seed-Code to develop small tools is also fast and efficient. We tried to input "Generate a pet diary App", and the model automatically built the App framework and designed the components.
In just a few minutes, the model delivered a ready - to - use finished product. Everything from the login interface to the upload of images and text works normally.
During the development process, we observed that Doubao-Seed-Code follows the logic of "plan first, then develop" and uses its in - depth thinking ability to self - analyze and optimize the generated results. When the user's requirements are not clear enough, the model can also actively sort out the requirements and even ask questions to obtain more information.
These features become an important foundation for the deployment of Doubao-Seed-Code in real production environments.
02. Can write both front - end and back - end code, and can modify production - level code libraries with a large context
Actually, the capabilities of Doubao-Seed-Code are far more than just creating interesting small programs or web pages. It can fully handle complex front - line development tasks.
In order to create a website that better meets real - world usage requirements, developers usually explain design details, interaction logic, and even technical constraints in great detail in the prompt words. This kind of "detailed instruction" poses higher requirements for the model: whether it can truly understand the user's intention and execute stably in complex tasks becomes the key to testing the model's strength.
In the actual test, we sent a very long prompt word to Doubao-Seed-Code at once, hoping that the model would create a prototype of an open - source project sharing website. The prompt word clearly specified specific components such as the top navigation bar, theme display area, and filtering tools, and also put forward detailed requirements for the design style.
The model not only accurately restored the design in the prompt word but also generated a directly interactive front - end page. The page layout is clear, the interaction logic is reasonable, and the overall style is highly consistent with the "technological sense" required.
In addition to prototype development, finding bugs is another major use of programming models. However, in a production environment, truly letting a large model modify code still carries the risk of introducing new errors, logical deviations, or security vulnerabilities.
Surprisingly, Doubao-Seed-Code not only has a native 256K context, which can search for and solve bugs in a large code library, but also shows good complex code repair ability and awareness of operating specifications.
We gave Doubao-Seed-Code a manually written Python file containing bugs and related folders to handle. It can first accurately locate the problem and notice various error reports and risks.
Doubao-Seed-Code repairs code in Claude Code
When repairing code, Doubao-Seed-Code adopts a step - by - step and incremental strategy - it verifies immediately after each modification to ensure that the changes are implemented.
More commendably, Doubao-Seed-Code is not limited to syntax repair. It can understand the program's logic and business requirements, actively improve exception handling and input verification, making the program more secure and reliable. Through continuous self - checking and iteration, it can discover potential problems left by previous repairs and further optimize them.
After challenging front - end design and bug repair, we also tried to let Doubao-Seed-Code go deep into the back - end database - this will further challenge the model's ability boundary.
Database tables, fields, relationships, and constraint rules are more abstract than front - end page elements. The model needs to understand the dependencies and functions between different fields. When designing database operations, how to ensure data consistency, avoid conflicts and redundancies is a challenge that tests the model's reasoning ability.
At this time, the planning ability of Doubao-Seed-Code comes in very handy. We can see that the system it creates has a clear structure, meeting the requirements for scalability and security in real production scenarios.
Of course, in such complex tasks, Doubao-Seed-Code can hardly complete all the work at once. When various bugs appear, we only need to provide the relevant code and error information to the model, and supplement it with a simple description, and the model can make further modifications and finally iterate out a usable database system.
We can also add complex functions, such as permission management, on the existing basis in the same way. After these functions are added, it can already meet the actual needs of some production environments.
03. Uncover the training formula of the Doubao programming model, and train a SOTA model with pure reinforcement learning
How was such a programming model, which is also very competitive in real production scenarios, developed? Volcengine said that in order to train Doubao-Seed-Code, they built a large - scale Agent reinforcement learning training system.
Data is an important source of intelligence. During the training process, Doubao-Seed-Code uses a huge training dataset covering 100,000 container images and provides an end - to - end sandbox environment for evaluation to ensure that the model can handle diverse and complex programming scenarios.
The model's training system has the ability to handle tens of thousands of concurrent sandbox sessions and relies on a cluster of thousands of GPUs to achieve ultra - large - scale parallel training, enabling the model to learn stably under high load.
Through reinforcement learning optimization, Doubao-Seed-Code learns directly from the feedback of the task sandbox. Based on this system, the model does not require cold - start data for distillation or annotation, and the training path is simple and efficient.
Downstream indicators show that the model performs well on the Multi - SWE - Bench and SWE - Bench - Verified datasets, showing a stable upward trend and demonstrating its generalization ability.
