OpenAI will hold its annual developer conference DevDay 2026 in San Francisco on September 29 local time, with the keynote scheduled to begin at 10 a.m. Pacific Time (1 a.m. Beijing Time on September 30). As the event approaches, OpenAI has ramped up its hype, teasing "20+ launches" for the day; at the same time, the company's next-generation model GPT-6.1 Astra, originally planned for an October release, has been cancelled because internal safety testing did not meet requirements. On one hand, there is an aggressive posture of密集 product launches; on the other, the next-generation model has been put on hold due to safety issues. OpenAI is at a critical juncture where product expansion and safety challenges coexist.

OpenAI posted a brief teaser on X on Monday, saying "1 day, 20+ launches," with the caption "Get ready." The company did not disclose what exactly these more than 20 launches include, nor how many are entirely new products versus feature updates, model upgrades, and developer tools.

The "Agent Wars" Begin as Meta Executive Takes a Long-Distance Jab

OpenAI's密集 teasers quickly drew the attention of competitors. Meta (META.US) Chief AI Scientist Alexandr Wang posted a "Star Wars"-style image on X, showing a fleet of X-wing-like spacecraft flying over a planet and approaching another. The distant planet features a subtle anthropomorphic expression, which observers believe bears a resemblance to the tilted bracket brand logo used in OpenAI's teaser. An X user subsequently commented that "the agent wars have already begun," pairing the remark with an image of legendary Jedi Master Yoda.

Behind this long-distance exchange is a shift in the focus of AI competition from chatbots to agents capable of completing tasks autonomously. Model makers are no longer just competing over the accuracy of answering questions; they are beginning to vie for AI's ability to execute tasks, call tools, operate computers, and connect to third-party services. Meta's new Muse AI agent launched last week has become a market focal point. According to TechCrunch citing estimates from Sensor Tower, since its release on September 8, Muse downloads have exceeded 3.4 million within weeks. Meta subsequently further expanded Muse's capabilities, adding computer-use functions and connectors, and plans to integrate it with AI glasses. Meanwhile, agent AI startup Instinct announced on Monday that it had completed $1 billion in funding at a valuation of $10 billion, roughly four times higher than before. The company is developing AI agents capable of autonomously handling daily tasks, including booking services and phone calls.

Astra's Delay Highlights Safety Challenges Facing Agent Development

The cancellation of GPT-6.1 Astra is particularly noteworthy because it was originally an important model in OpenAI's move toward a higher degree of autonomy. According to reports, Astra was originally planned for an October release and was expected to be used in ChatGPT and Codex, with a focus on handling more complex tasks requiring less human intervention. Reports said OpenAI's internal testing found that the model did not meet the company's safety and alignment standards in areas such as following instructions, authorization scope, and accurately disclosing its own actions, leading to the decision to cancel the planned release. According to reports, Saachi Jain, head of OpenAI's safety systems, said that although the model showed improvements in areas such as reducing "model laziness," it still fell short of company standards in adhering to task scope, authorization requirements, and accurately explaining to users the work it performed. This also gives OpenAI's DevDay product offensive another layer of meaning: as AI agents are given increasingly stronger autonomous execution capabilities, how to keep models within user authorization while granting them greater abilities will become a problem the industry must solve. Yet competition has not slowed because of this. Anthropic launched Claude Opus 5.5 on September 22, positioned for complex agent work; on September 28, it launched Claude Sonnet 5.5, targeting faster, lower-cost daily tasks. AI companies are accelerating their push into the agent market from different directions.

From "Strongest Model" to User Distribution, OpenAI Seeks a New Moat

While OpenAI accelerates its launches, investors are also beginning to focus on its competitive advantages beyond models. On Monday, David George, a general partner at OpenAI's major investor a16z, published a blog post saying that OpenAI's long-term advantage may depend less on whether it has the "best single model" and more on whether it can create new user behaviors and distribute AI products to a massive user base. George noted that as the cost of switching models decreases, AI companies are increasingly competing around price and performance. In his view, creating new types of user behavior and distributing AI products to large numbers of users will become more important competitive levers. He said OpenAI has already continued to drive new ways of using AI through features such as ChatGPT's chat interface, reasoning, tool calling, and computer use, while its huge consumer and enterprise user base can also form a continuous feedback loop. At its 2025 DevDay, OpenAI previously announced a series of developer products and features including AgentKit, Apps in ChatGPT, Sora 2 in the API, GPT-5 Pro, Codex upgrades, gpt-realtime-mini, and gpt-image-1-mini. As this year's DevDay approaches, the market's focus is no longer just on how many new products OpenAI will release, but on whether these products can further push AI from "answering questions" to "completing tasks for users" and build a stronger developer and user ecosystem in the agent competition. According to Nasdaq Private Market data, OpenAI's latest private market valuation is $910.26 billion.