OpenAI Codex Enters Mobile

On May 15, 2026, OpenAI released an update that sparked significant discussion in the tech community. Codex has officially entered the ChatGPT mobile app. Now, users can check what tasks AI completed for them while waiting for the subway or approve long-pending decisions. The AI continues to work in the background, allowing users to focus on other activities.

This update is not just a new toy for programmers; anyone can use it to accomplish tasks like writing reports, organizing data, processing documents, and automating workflows. AI is now truly available on demand.

The Ongoing Agent War

To understand why Codex is exciting, we need to clarify what it does. Traditionally, using AI involved a question-and-answer relationship: you ask a question in ChatGPT, it responds, and you execute the instructions. Codex changes this dynamic. It acts as an agent that takes on tasks and completes them independently.

With the mobile launch, Codex can operate without the user being tethered to a computer, working while they sleep, commute, or attend meetings.

However, Codex is not the first to think of this. Other tools like Anthropic’s Claude Code and OpenClaw, which has gained popularity in the Chinese internet as “龙虾” also offer similar capabilities. OpenClaw was among the first to emerge, being open-source and free, capable of integrating with any large model, and featuring persistent memory that improves with use. It can be embedded into messaging apps like WeChat and DingTalk, receiving commands through these platforms and even setting up scheduled tasks.

Despite its advanced functionality, OpenClaw has a fundamental issue: it is a geek’s toy. Users must deploy it themselves, configure APIs, and troubleshoot bugs, which often leads to abandonment after reading the tutorials. There have been real security incidents where users faced issues like bulk email deletions and unauthorized credit card charges.

Claude Code and Codex, on the other hand, operate beyond this boundary, covering a larger market. Claude Code entered developers’ minds earlier this year, introducing remote control features in February, allowing developers to manage sessions on their computers via QR codes. However, it has not gained as much traction as Codex, primarily due to its high cost and the command-line interface that deters average users.

Codex differentiates itself with lower pricing, a smoother product experience, and the recent mobile launch, making AI work accessible to everyone. The relationship among these tools can be summarized as follows: OpenClaw validated the demand, Claude Code educated the market, and Codex is now reaping the benefits.

Codex’s recent actions signal a clear message: the agent battlefield is expanding from developers to the general public. The direction is clear: lower prices, lower barriers, and always available.

Domestic Competitors Need to Wake Up

In recent months, the Chinese tech industry has seen a surge in interest in “raising lobsters,” with major companies like Alibaba, Tencent, ByteDance, and Baidu launching their own OpenClaw-integrated products. Local governments have also entered the fray, offering subsidies of up to 5 million yuan. The media has been flooded with buzzwords like “AI digital employees,” “one-person companies,” and “productivity revolutions.”

However, a crucial question remains unanswered: can these products actually perform tasks? Most domestic “lobster” products follow a similar path: they build on the OpenClaw framework, integrate their large models, and connect to existing traffic channels like WeChat and DingTalk. This approach is the fastest, cheapest, and easiest to market.

Yet, OpenClaw is just a framework; the true capability of an agent depends on the underlying model. Tasks assigned to agents require high reasoning ability, instruction comprehension, and error handling. Codex is backed by OpenAI’s latest model, while Claude Code relies on Anthropic’s Claude.

Despite significant progress in domestic large models over the past two years, there remains a gap in the quality of completing complex tasks. Short-term engineering solutions cannot bridge this gap. As a result, users who eagerly adopt these “lobster” products often find themselves with incomplete results or processes that require repeated corrections.

Once the novelty wears off, user retention becomes the real challenge. The commercial logic behind these products relies on agents driving token consumption and cloud service revenue. If agents cannot perform well, users will not continue to use them.

Without sustained usage, token consumption becomes irrelevant, and the potential for cloud services diminishes. Furthermore, Codex’s recent actions have raised the benchmark. Users may have previously been unaware of the capabilities of agent tools and settled for less, but with Codex’s accessibility, transparent pricing, and smooth experience, user expectations have shifted.

Once expectations are elevated, users will quickly lose patience with products that are functional but not user-friendly.

How Long Can the Domestic “Lobster Dream” Last?

So, how will this battle ultimately unfold? One detail is noteworthy: on the same day Codex launched on mobile, OpenAI disclosed that over 4 million people are using Codex weekly. This indicates that the adoption of agent tools is happening faster than most anticipated, and the window of opportunity is closing.

For domestic manufacturers, the most significant threat is not today. Currently, Codex has not fully integrated into the Chinese language environment or deeply connected with domestic office ecosystems, and localization efforts are nearly non-existent. However, the question remains: how long can these barriers hold?

As people around them start using Codex to complete tasks, the comparative disadvantage of local products will shift from data to reputation, and from reputation to user attrition.

More critically, the essence of this competition is not a product battle but a model battle. The accumulation of model capabilities requires time and resources, with no shortcuts.

Domestic manufacturers face two paths: one is to continue catching up with general agent capabilities and compete directly. This path is the hardest, requiring genuine investment in models rather than just ecosystem packaging. However, if successful, the rewards could be substantial. The other path is to abandon general capabilities and focus on vertical markets like healthcare, law, manufacturing, and government affairs, where data security, local deployment, and industry knowledge are crucial, presenting an advantage for local firms.

Regardless of the path chosen, one thing is clear: the era of attracting users solely through the “lobster” concept is over.