Are 2026 AI programming tools shells? Detailed explanation of the real and fake domestic models of Cursor, Windsurf and TRAE.

Cursor, Windsurf, and Byte TRAE, these three mainstream AI programming tools have been repeatedly asked the same question recently: are they their own products, or are they the shell of domestic or foreign large models. This matter seems to be just technical gossip, but it involves user payment psychology, corporate compliance, and data flow. This article will not go into specific figures, but will only explain clearly the controversy of "shelling" and how to understand each of the three tools.

What is the underlying model of Cursor?

Cursor is developed by the Anysphere team, which has long publicly claimed to use flagship models from OpenAI and Anthropic. Users can choose GPT series, Claude series, Gemini series and other public models in the settings. This in itself is not controversial.

What really caused the discussion was Cursor's "fast" path. This path exhibits different latencies and answering styles during peak periods than direct connections to Anthropic or OpenAI during the same period. Some people in the community suspect that it dynamically routes to other cheaper or faster inference backends during certain periods, but Anysphere did not admit it directly, only saying that "it will use the most appropriate model to fulfill the request." This wording itself gives routing space, but it cannot be called "shelling".

Shelling and routing are not the same thing

Many people mix the two. "Shelling" in the strict sense means that the product has no self-development capabilities and only forwards requests to third-party APIs and includes a layer of UI. Cursor obviously does not belong to this category. It has a large number of its own implementations in the engineering layers such as editor integration, context retrieval, agent scheduling, and completion indexing.

Routing is another phenomenon. An AI tool may switch the actual call behind the same "Claude" button to different models based on the request type, user profile, and current load. This is not a shell in the technical sense, but if the user thinks they are getting model A when they pay, but actually get model B, it involves the issue of information disclosure.

Windsurf ownership changes

Windsurf, formerly Codeium, started with automatic code completion, and later shifted its focus to agent mode. In 2025, there was news that it was acquired by OpenAI, but public information showed that the transaction was not completed in the end, and Windsurf's core team was later absorbed by another large American company. In other words, the ownership and route of the Windsurf brand have undergone significant fluctuations in the past year. The latest situation shall be subject to the official announcement.

For users, the impact of this ownership change is: model selection, pricing methods, and corporate compliance commitments may all be adjusted with changes in the parent company. If you are evaluating it, it is best to look directly at the model list and price on its current official website, rather than relying on your impressions half a year ago.

Does Byte TRAE use its own Doubao?

Byte's TRAE is Byte's self-developed AI programming IDE. It is positioned against Cursor and is gaining popularity in the domestic AI development circle. Judging from public information, the default inference backend of the domestic version of TRAE is Byte's own Doubao series of large models, while the overseas version is more inclined to accept the public models of Anthropic and OpenAI. This kind of "different backends at home and abroad" is mainly due to compliance needs. Domestic commercialization to provide foreign model services requires special channels and cannot be circumvented.

Does it count as "setting one's own shell"? From the perspective of product caliber, Byte does not package TRAE as "the world's most powerful model driver", but positions it as "the entrance to AI programming in the Byte ecosystem". Therefore, using Doubao is logical and not misleading. What really needs to be asked is whether Doubao is sufficient for coding scenarios. This is a matter of opinion.

国产模型在代码场景的真实位置

By 2026, the overall level of domestic code models will be significantly improved compared to a year ago. The experience of Kimi, GLM, Doubao, and Alibaba Qwen in daily tasks such as code completion and function generation is close to the mainstream closed-source model, but there is still a gap in long context reconstruction and complex agent chains. How big is the specific gap? The public list scores of each company change frequently. It is safer not to cite the specific scores. It can only be said that in terms of Chinese annotations, domestic framework understanding, and local tool chain calls, domestic models have their own home court.

This means that even if an AI programming tool is routed to a domestic model at certain times, it will be difficult for users to detect it at the level of daily completion and small function generation. Differences mainly appear in large refactorings and complex agent tasks.

Common sense ranges for price and value for money

I won’t quote specific package numbers that may be wrong, just talk about general knowledge. Cursor's personal subscription is roughly in the same price range as ChatGPT Plus and Claude Pro, which is about twenty dollars per month. The enterprise version is more expensive. Please refer to the official price list for details. Windsurf's personal subscription tier is close to Cursor, but it may be adjusted at different times. The domestic version of TRAE is priced significantly lower than the US dollar subscription tool, which is a common structure in the Chinese market. Claude Code itself is priced according to Anthropic's Pro/Max subscription system, and there is no independent package of "tens of dollars per month with millions of tokens" circulated in some community posts.

If you want to choose, remember this: the difference between Pro gears is actually much smaller than the difference caused by usage habits. The output of a developer who knows how to use prompts using a cheap tool may be higher than that of a casual clicker using a flagship subscription.

How to tell which model you are using

No complicated tools are needed, and there are several simple ways to make rough judgments. Asking the model "Who developed you?" can sometimes reveal the identity, but more and more models are trained not to reveal it proactively. You can make fuzzy judgments based on how it handles sensitive topics in Chinese. The domestic model tends to be conservative in response, while the overseas model is more direct. Looking at the output style can also give some clues. Different families of models have preferences in Markdown use, list format, and sentence structure. None of these methods are 100% accurate, they just give a level of confidence. If you are really concerned, the safest thing is to look at the official model disclosure document and the usage details in the account.

To whom is this matter most important?

For enterprise users, what they care about is not which model is cool, but whether code requests will pass through undisclosed inference nodes. If engineering code involving trade secrets is routed to inference services that are not in the procurement compliance list, there will be data governance risks. This is the fundamental reason why many large enterprises have explicitly banned or restricted the use of AI programming tools. It has little to do with whether the model is useful or not.

The impact on individual developers is relatively small, but if you spend a subscription fee but don’t get the model you expected, the experience will indeed be uncomfortable. A more practical way to deal with this situation is to see if the tool provides a "lock model" option, turn off the "automatic/fast" path, and fix it to a public model.

How to choose the AI ​​programming tools of 2026

If delay and price are the primary considerations for domestic developers, the domestic version of Byte TRAE or direct access to the official APIs of Doubao, Kimi, and Zhipu are convenient options. Overseas developers or people who are deeply dependent on the Claude/GPT ecosystem can choose one of Cursor, Claude Code, and Windsurf. The main differences lie in the details of the agent mode and IDE integration. Enterprises with high security requirements are best to smuggle deployment routes and use open source IDE plug-ins such as Cline, Continue, and Aider to connect to their own reviewed API endpoints. Students and beginners can use the free tier to experience it first, so they don’t have to buy Pro right away.

FAQ

Does Cursor really secretly use domestic models?

There is no official recognition. There have been doubts about latency and style drift in the community, but Anysphere's public statement is that "it will use the most appropriate model to complete the request", which gives room for routing. If you mind, you can turn off the fast path in the settings and lock the specific Claude or GPT model, so that the request path will be much clearer.

What is the difference between TRAE overseas version and domestic version?

The main differences are in the default model and compliance channels. The domestic version tends to use Byte's own Doubao series and the domestic ecosystem connected with it, while the overseas version tends to use public models such as Anthropic and OpenAI. The specific price and functions of each version will be adjusted as the version is updated, please refer to the current page of TRAE official website. The two versions have different regional access experiences, which will be affected by VPN and IP ownership.

Is Windsurf still worth using now?

It's worth it, but you must first see clearly its current ownership and model route. Windsurf has experienced acquisition rumors, team changes and direction adjustments in the past year. If you came here for a specific model before, go to its current official website to confirm the default engine and price before making a decision.

Is it cheaper to use VS Code and Cline or Aider yourself?

It is more economical, but it takes time to configure. Open source agent and your own API key can achieve most of the daily experience of mainstream IDE tools, and the cost is mainly API token consumption. The price is the lack of indexing, completion models and team management that come with specialized tools, and the context recall of large monorepo may be inferior. This combination is sufficient for personal scripts and small projects, but commercial tools are recommended for large enterprise-level code bases.

Is it possible that the supply of AI programming tools will be cut off by the country?

U.S. export controls do impose bans on some countries, and China is currently not on the strictest embargo list. However, OpenAI and Anthropic have always used third-party agents or local compliance channels to directly provide commercial API services to mainland China, and the delay and stability are worse than direct connections. The safe way is not to put all your engineering links on a single overseas model, but keep a domestic model as a backup. Just choose any one of Kimi, Zhipu, Doubao, and Qwen as a backup.

Source of inspiration: Ruan Yifeng's "Kimi Cursor's Suspense" https://www.ruanyifeng.com/blog/2026/03/kimi-cursor.html