On November 6, 2023, OpenAI held its first developer conference. The announcements covered considerable ground. GPT-4 Turbo, with a hundred and twenty-eight thousand token context window and significantly lower per-token pricing. Custom GPTs, which let any user create a specialised version of ChatGPT with custom instructions and access to specific tools. The Assistants API, which gave developers a way to build agent-like applications with persistent state, tool use, and code execution. A GPT Store, planned for the following year, that would let users discover custom GPTs created by others.
Together the announcements outlined a clear platform ambition. OpenAI was not content to be the API provider behind applications built by other companies. It intended to become the platform on which a new category of AI-powered applications would be built and discovered.
The pricing changes mattered immediately. The cost of running applications built on GPT-4 dropped substantially with the move to GPT-4 Turbo. Use cases that had been economically marginal at previous prices became viable. The competitive pressure on alternative providers, including Anthropic, Google, and the major open source efforts, ratcheted up.
The custom GPTs announcement landed differently with different audiences. For users, the appeal was obvious. The ability to create a specialised assistant for a particular domain, embed instructions and reference materials, and use it without writing code, was a meaningful expansion of what ChatGPT could be used for. For startups that had been building thin wrappers around the API, the announcement was uncomfortable. Many product categories that had been treated as differentiated startups were now potentially replicable in minutes by anyone with a ChatGPT Plus subscription.
The Assistants API was the most technically significant of the announcements. Building agent-like applications had been possible with previous APIs but required significant infrastructure. The Assistants API encapsulated common patterns and provided primitives that made building complex AI applications considerably faster.
What the day represented strategically was OpenAI placing itself at a different layer of the stack than it had occupied previously. The company was no longer just selling intelligence as a service. It was attempting to become the operating system for AI-native applications. Whether that ambition would be realised, partially realised, or actively contested by competitors and customers, would unfold over the following years.