OpenAI product strategy

Aside from investment mega-deals with NVIDIA, AMD, Oracle and Broadcom, which took most of the mainstream media attention, OpenAI has this year made a number of product announcements and launches that shed light on their product strategy going forward.

I don't have any inside knowledge, but here are a few thoughts on what I think is going on.

Platform

Initially, OpenAI was all about building the most powerful AI models and lettings other use it via an API. With the appearance of strong competitors, all of which used the same API for compatibility reasons, the LLMs themselves became a commodity. A very expensive one to build, yet with a downward pressure on the price you can charge for it.

They countered it by layering new APIs such as Assistants and later Responses API, to prevent trivial switching away from the OpenAI platform. Neither of those changed much, mostly because they didn't provide a lot of extra value on top of the commodity Chat Completions API.

This was followed up with Agents SDK, which was a better abstraction than Assistants, and recently with the Agent Builder and ChatKit, creating a unified AgentKit toolkit. This provides more powerful abstractions, making it easier to build AI-powered agents and chatbots. At the same time, they integrate tightly with the rest of OpenAI platform, making it much harder to switch to a different provider.

ChatGPT

On the end-user products side, ChatGPT was an technology demo turned unexpected success they didn't really know what to do with. An initial attempt to build a platform out of it with custom GPTs came to nought - it was poorly thought out, poorly implemented, and poorly received.

Monetization proved to be a problem, with many users sticking to a free version and power users potentially using it more than they pay for it (negative margins). This was exacerbated by the appearance of new, larger, reasoning models burning through a lot of compute to answer a simple query, and necessitating the price increase. A new, 10x more expensive Pro plan is still losing money.

Naturally, they also segmented business and enterprise users with standard security and compliance features like SOC2 certification, SSO, MFA and other features specifically required by those types of customers.

Competition

However, with white-hot competition (Anthropic for models, Gemini for business and enterprise, Perplexity for search), there was a clear risk for ChatGPT to slide down to "one among many" AI tools. In particular, Anthropic blind-sided OpenAI in two areas:

First, Claude Code has won the developer hearts practically overnight. OpenAI countered after a few months with special-purpose Codex AI model, then Codex CLI and Codex Cloud devtools, regaining the balance somewhat.

Secondly, with MCP support becoming wider, it was easy for anyone to connect to Claude, building integrations "for free" for Anthropic. This is useful from a user perspective, but massive from a business perspective. OpenAI had no choice but to join the competition and support MCP in ChatGPT with their Data Connectors.

Google, pushing now competent Gemini models everywhere, including in Chrome and AI search overviews, is possibly the biggest challenge. Why do you even need ChatGPT if Google has an AI to answer your search queries, and is building one right inside Chrome, the most used browser on the planet?

Atlas and Apps SDK

I believe ChatGPT Atlas and Apps SDK are two-pronged answer to this challenge.

Atlas is a new browser powered by Chromium, the same engine that powers Chrome itself, Atkas tightly integrates ChatGPT within the browser experience, attacking both Chrome and Google Search. Who needs to google something if you can just ask your browser? This also addresses the competition from Perplexity Comet, also an AI browser.

The Apps SDK is a way to bring other apps directly into ChatGPT interface. In contrast to MCP, which gives ChatGPT some new abilities without changing the user experience, Apps SDK provides a way for the 3rd party apps themselves to run "within" ChatGPT. The most apt analogy is with Facebook Apps (remember when those were all the rage)? If successful, this strategy would turn ChatGPT into a main user interface for computer interaction in general, one-upping Chrome and google as the main user interface for "the internet".

Atlas vs Google

Atlas seems to attack Google from below, in a classic case of Clayton Christensen's "Innovator's Dilemma", but I don't think that will work.

First, Google doesn't actually make money from Chrome itself. Adding Gemini deeply into Chrome costs them nothing in terms of lost revenue, and they're already doing exactly that. The gap between current Chrome AI features and what Atlas or Perplexity's Comet offer is actually pretty small.

More fundamentally, I don't see how Atlas can actually undercut Google in the way classic disruptive technologies do:

The advertising problem: The moment OpenAI adds advertising to Atlas to monetize free users, they become just another Google competitor, except Google has overwhelming technological and business advantages in the ads space that took decades to build.

The paid user niche: If they focus purely on paid subscribers, that market is too small to pose a serious near-term threat to Google.

The cost structure problem: Once you download Chrome, you cost Google essentially nothing. But every query in Atlas costs OpenAI inference budget. The ongoing costs of maintaining a large free user base scale rapidly and unsustainably. Google has far more cash to defend their position than OpenAI has to give away their product for free.

The shareholder calculus: Google has boatloads of cash and such a dominant position that shareholders would not be upset if it absorbs short to medium-term costs to protect its position. In fact, they would probably be far more upset if management didn't respond to threats like Atlas.

Apps in ChatGPT

Allowing apps to integrate with ChatGPT seems like a more promising strategy, if OpenAI can bootstrap a vibrant ecosystem around it. They can move faster than Google, and are further along than Anthropic or other competitors. But they will have to execute well, and quickly, because fundamentally it is still easy to copy.