Furcht nicht vor der Unterbrechung von Claude, hier kommt Doubao's Programmier-Modell. Es kann in 5 Minuten eine Kopie von "Minecraft" erstellen und kostet nur 0,2 Yuan.

Doubao's first programming model is here!

According to a report from Zhidx on November 11th, today, Volcengine, a cloud and AI service platform under ByteDance, has unveiled the first programming model from the Doubao large - model family – Doubao - Seed - Code. This is a programming model that is specifically optimized for agentic coding tasks and has achieved a breakthrough in cost - effectiveness.

In terms of performance, Doubao - Seed - Code has achieved better results than Chinese models such as DeepSeek - V3.1, Kimi - K2, and GLM - 4.6 in several industry - standard programming test sets. Overall, it has the second - best performance, only behind the current top model in the field of AI programming – Claude Sonnet 4.5. In addition, Doubao - Seed - Code has a built - in 256K context, which is higher than the 200K context of Claude Sonnet 4.5.

Beyond the rankings, Doubao - Seed - Code also attaches great importance to implementation in real - world programming scenarios. Thanks to its special optimization for common development tools, users of Claude Code, Trae, or veCLI can use it without any problems and achieve stable output results.

At the same time, Doubao - Seed - Code is the first Chinese programming model that supports visual understanding capabilities. It can generate code based on UI designs, screenshots, or hand - drawn sketches, or perform visual comparisons of generated pages and automatically make style corrections and bug fixes, significantly improving the efficiency of frontend development.

It's worth noting that today, the Chinese version of ByteDance's AI - native IDE product Trae has integrated the Doubao - Seed - Code model. The combination of Trae + Doubao - Seed - Code has topped the authoritative programming benchmark SWE - Bench - Verified, thus creating an ecosystem closure of the model and the tool.

In combination with the price, Doubao - Seed - Code becomes even more powerful. This model uses a tiered pricing model. In the most commonly used input range of 0 - 32K, the input price is 1.20 yuan per million tokens, and the output price is 8.00 yuan per million tokens. After using the full transparent cache, the usage cost of the model can be reduced by 80%, and the total usage cost drops by 62.7%.

In our practical tests, we asked Doubao - Seed - Code to reproduce the classic game "Minecraft". The cost was less than 0.2 yuan, and both the graphic design and the gaming experience were similar to the original, and it was directly playable. With the same number of tokens (in the range of 0 - 32K), Claude Sonnet 4.5 has already incurred costs of over 3 yuan for the same task.

Recently, Zhidx has intensively tested the effectiveness of Doubao - Seed - Code in real - world programming scenarios for the first time. It can not only independently create a development plan and quickly build frontend websites but also deeply intervene in the database and make changes. In case of errors, it automatically corrects them, adds comments, and optimizes the structure. From now on, Doubao - Seed - Code is no longer just a "code - writing machine" but a development partner that can think and create together with humans.

01. Seamless integration with Claude Code – you can reproduce "Douyin" with a screenshot

Tool compatibility has always been one of the main factors affecting the adoption of AI programming models. This time, Doubao - Seed - Code has invested a lot of effort in adapting to tools. Doubao - Seed - Code is natively compatible with the Anthropic API and can be directly used in Claude Code without any conversion. This means that many developers who are familiar with Claude Code can switch to Doubao - Seed - Code with almost no learning curve.

The Volcengine Ark platform also provides a detailed guide for calling the Doubao - Seed - Code API. Even absolute beginners can simply follow the guide and test the new model.

Before we intensively test the development capabilities of Doubao - Seed - Code, let's give you a few "appetizers" first.

Ball bouncing has almost become a standard test for large models. The ball - bouncing created by Doubao - Seed - Code not only conforms to the laws of physics but is also very smooth. At the same time, the model has autonomously decided to add a new function: when you click on the ball, you can change its force application, so that the ball's bouncing is no longer just an endless repetition.

We have also tested the model's ability to develop websites based on screenshots. After uploading a screenshot, Doubao - Seed - Code can analyze the page layout and visual features and then gradually build the core components of the website. Before sending the result to the user, the model also conducts a function test and then delivers the finished product.

In daily application scenarios, developing small tools with Doubao - Seed - Code is also fast and efficient. For example, we entered "Create a pet diary app", and the model automatically created the app framework and designed the components.

Within a few minutes, the model delivered a directly usable product. Everything from the login screen to uploading images and texts worked perfectly.

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 analyze and optimize the generated results by itself. If the user requirements are not clear enough, the model can organize the requirements by itself and even ask questions to obtain more information.

These characteristics form an important foundation for the implementation of Doubao - Seed - Code in real - world production environments.

02. Can write both frontend and backend code, and the large context enables changes to production code libraries

In fact, the capabilities of Doubao - Seed - Code go far beyond creating interesting mini - apps or websites. It can handle fully complex development tasks in practice.

To create a website that better meets the actual usage requirements, developers usually specify in detail the design details, interaction logic, and even technical limitations in the instructions. These "precise instructions" place higher requirements on the model: whether it really understands the user's intention and can work stably in complex tasks is the key to testing the strength of the model.

In our practical tests, we sent Doubao - Seed - Code an extremely long instruction and asked it to create a prototype of a website for open - source project release. The instruction precisely defined specific components such as the top navigation bar, the theme display area, and the filter tools and also put forward detailed requirements for the design style.

The model not only accurately reproduced the design described in the instruction but also generated a directly interactive frontend page. The page layout is clear, the interaction logic is reasonable, and the overall design style highly conforms to the required "technological style".

Besides prototype development, finding bugs is also an important function of programming models. However, in the production environment, there is still a risk that using large models to change code may cause new errors, logical deviations, or security vulnerabilities.

Surprisingly, Doubao - Seed - Code not only has a built - in 256K context, which enables it to search for bug solutions in large code libraries, but also shows good ability to repair complex code and awareness of operating guidelines.

We gave Doubao - Seed - Code a manually written Python file with bugs and the associated folders. It can first accurately locate the problems and pay attention to all errors 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 immediately checks whether each change is successful after making it.

Even more remarkable is that Doubao - Seed - Code is not limited to syntax repair. It can understand the program logic and business requirements and actively improve error handling and input validation to make the program safer and more reliable. Through continuous self - checking and iteration, it can discover and further optimize potential problems left over from previous repairs.

After testing the frontend development and bug repair, we also tried to integrate Doubao - Seed - Code into the backend database – this will further test the capabilities of the model.

Database tables, fields, relationships, and constraint rules are more abstract compared to frontend page elements. The model must understand the dependencies and functions between different fields. When designing database operations, it is a challenge for the model's inference ability to ensure data consistency and avoid conflicts and redundancies.

Here, the planning ability of Doubao - Seed - Code comes into play. You can see that the system structure created by it is clear and meets the requirements for scalability and security in real - world production environments.

Naturally, it is also difficult for Doubao - Seed - Code to complete all the work at once in such complex tasks. If there are various bugs, we only need to send the relevant code and error messages to the model and add a simple description. Then the model can make further changes and finally deliver a usable...