Last week, many users received an unexpected update from Microsoft: "GPT Builder is being retired" The announcement stated that support for Copilot GPT would end on July 10, and all GPTs created on the platform, along with related data, would be deleted within four days.
This sudden change comes just three months after GPT Builder was officially launched as part of the Copilot Pro service. Subscribers paying $20 per month could create custom GPT chatbots for personal use or sharing based on specific needs.
Regarding the abrupt shutdown, Microsoft explained: "We are continuing to evaluate our strategy for consumer Copilot extensibility and are prioritizing core product experiences, while remaining committed to developer opportunities. To this end, we are shifting our focus on GPTs to Commercial and Enterprise scenarios and are stopping GPT efforts in consumer Copilot. "
A quick fall
In short, after a few months of experimentation, Microsoft concluded that GPT Builder is not a viable product for the consumer market. While the company continues to push AI projects like Copilot, support will only be given when they have real commercial significance.
Similarly awkward is the fate of the Copilot GPTs' "body," the ChatGPT-based GPTs, and the GPT Store, which operate under the same product logic.
As GPT-4, Sora, Voice Engine, and GPT-4o spark waves of discussions, OpenAI's toolkits have produced numerous hits, while GPTs have gradually faded into the background, rarely mentioned. However, people still remember its grand debut—a potentially game-changing product introduced in Silicon Valley during the last winter.
On November 6 last year, at the OpenAI developer conference, Sam Altman introduced the concept of GPTs to the world. Users could easily customize their personalized GPTs using natural language commands and uploaded knowledge bases without any programming knowledge.
GPTs featured multimodal capabilities, supported external data source extensions and API integration, and could be published on the upcoming GPT Store for community sharing and profit. Altman confidently described the GPT Store's search, categorization, and recommendation mechanisms akin to the App Store and personally demonstrated how to create a fully functional "startup mentor" chatbot in minutes.
Suddenly, the technical barriers for ordinary people were removed, redefining the term "developer" and causing a significant shakeout among numerous AI startups. People saw OpenAI's ambition to emulate Apple and Google in building a comprehensive developer and user ecosystem. At that time, words like "revolutionary," "explosive," and "disruptive" were not yet common, but that launch event was described as "OpenAI's iPhone moment."
Unexpectedly, a dramatic coup disrupted the GPT Store's release schedule, delaying its launch from late November to January 10 for Plus users. Meanwhile, issues with using GPTs emerged.
A product that is going nowhere
With low technical requirements, GPTs quickly attracted a large global community of AI enthusiasts. By the end of January, over 3 million GPTs had been created. However, it became apparent that these custom GPTs, created with simple prompts, offered experiences not significantly different from direct interactions with GPT-4, often feeling like a "downgraded version": they struggled with knowledge bases, were easily distracted, hallucinated, and couldn't fully follow custom instructions. Essentially, they functioned as basic RAG platforms, limited to uploading ten files and lacking features like vector database calls. Their ease of creation also meant ease of replacement, leading to a flood of homogenous, low-quality GPTs with few truly excellent ones.
As a result, people preferred using a powerful, flexible, and general-purpose AI assistant for various complex needs. According to Sametime Web data, GPTs accounted for only 1.5% of ChatGPT's web traffic in February. Since the GPT Store's opening, weekly user traffic had stabilized.
More concerning were the security risks posed by some independent GPTs. Many users reported on X that they could easily extract original knowledge files from GPTs through a few conversations, posing a data leakage risk. For example, someone obtained a 2021 salary data spreadsheet for major U.S. tech companies from Levels.fyi's GPT. Another user extracted the source prompts for 16 official OpenAI GPTs. Ordinary users could replicate these GPTs by copying prompts and pairing them with knowledge base files.
Developers were also unhappy. Until the May 13 OpenAI spring release event, only paid users could access the GPT Store, significantly limiting GPTs' user growth potential. Additionally, the lack of user data analytics from OpenAI's backend made it difficult for creators to gain insights in a data-driven decision-making world, hampering their ability to attract new users.
Kirill Demochkin, founder of Hintloop, stated that among over 36,000 custom chatbots he analyzed, only 5% had 150 to 500 active users daily, while most attracted only one or two users daily.
New creators also faced visibility issues in the GPT Store. Although it was open to everyone, the store's current settings made it hard for new creators to gain attention. Each category only showcased OpenAI's official recommendations. The community called for OpenAI to adopt a big data push or discovery method like TikTok, giving new creators more opportunities to showcase themselves.
Furthermore, OpenAI's ambiguous monetization policy was a major issue. Despite Altman's promise at the developer conference that GPTs would generate actual revenue, no payment rules had been announced. Without a share of the profits, there was little incentive to build high-quality applications.
By late March, seeing GPTs on the verge of "cooling down," OpenAI urgently decided to collaborate with a small number of U.S. developers, offering rewards based on GPTs' usage. The goal was to "create a vibrant ecosystem where developers' creativity and influence are rewarded." However, OpenAI did not disclose specific reward details.
In stark contrast is the AI chat platform Poe, created by OpenAI board member Adam D'Angelo. With a similar creator economy model, Poe not only acted earlier but developed a clear and comprehensive monetization plan, extending its reach to 23 countries and regions.
On Poe, creators could earn revenue by setting chatbot paywalls and pricing per message, or by attracting new users, earning up to $50 per subscription incentive. Poe thoughtfully explained the revenue calculation method, how to set up payments, and tips for increasing conversions. Rumors had it that D'Angelo was furious at being sidelined by Altman, yet Poe's path seemed smoother while GPTs stumbled.
Despite attempts to revive interest by introducing a feature allowing direct mentions of GPTs in conversation boxes, making 3 million chatbots "work for you," the cumbersome process of finding, mentioning, selecting from a dropdown menu, and entering commands made it unlikely to gain traction. Last month, OpenAI announced free access for all users, but this seemed to have little effect.
Why didn't GPT Store become the App Store?
Bearing Altman and OpenAI's ambitions, from its sensational debut to its diminished presence, GPTs have not yet become what they envisioned.
Top researchers at OpenAI once suggested that LLMs could become an operating system, similar to Apple's iOS. However, early popular apps for iOS and Android included video games, navigation maps, social media, and browser tools—services with clear functions that were easy to use and understand. In contrast, early generative AI-driven applications were not as intuitive, and most people hadn't formed habits of using chatbots for tasks like creating charts, planning trips, or solving math problems.
Analysts also noted that the success of the App Store and Play Store was partly due to their integration with essential consumer devices like iPhones and Android phones, embracing a massive user base. While ChatGPT had amassed an impressive 180 million users, it remained a novelty for many. It's no wonder Altman hasn't given up on AI hardware plans, aiming to secure this ecological high ground.
However, the failure of GPTs and GPT Store does not impact OpenAI's revenue capabilities. According to The Information, OpenAI's annualized revenue doubled to $3.4 billion in the past six months. At the end of 2023, this figure was $1.6 billion, compared to $1 billion in the previous summer. Anthropic plans to achieve over $850 million in annualized revenue by the end of this year, while Canadian competitor Cohere's annualized revenue was only $22 million in April—OpenAI still leads its peers by a wide margin.
Of OpenAI's $3.4 billion, $3.2 billion comes from ChatGPT subscription services and software developers accessing its models via API, with Microsoft taking a share since these models are deployed on Microsoft's cloud. OpenAI also earns from Microsoft's sales of OpenAI models to its Azure cloud customers, generating approximately $200 million annually, only 20% of the revenue Microsoft derives from this business.
Now, with Microsoft discontinuing consumer Copilot GPTs to focus on enterprise-level applications, it aligns with OpenAI's business direction. This year, OpenAI COO Brad Lightcap publicly stated multiple times that the company is exploring more B2B models, including comprehensive custom services for enterprise models.
GPTs seem more like an experimental product of the AI era. They replaced the initial GPT plugins and created global buzz. Although they have now faded, it's possible that for every hundred GPTs that fall, a thousand other innovations will rise.