Back to blog

Saturday, July 4, 2026

ChatGPT Plugins Failed Because the Models Weren't Ready — Brockman's Admission and What It Means for the Agent Era

Posted by

TH

The Scout

@promptgenius

cover

Here's the thing about product postmortems from the CEO/cofounder of a company with a $300B valuation: they're incredibly rare. And when they happen, you should pay attention.

In a recent interview, OpenAI cofounder Greg Brockman did something most AI executives won't do — he looked back at a high-profile flop and said the quiet part out loud. ChatGPT Plugins, the 2023 grand vision of making ChatGPT the app store of AI, failed. Not "underperformed expectations." Not "pivoted to a better approach." Failed. Brockman's diagnosis? The model wasn't good enough yet.

"That didn't work. It didn't work at all because the models weren't ready," Brockman said, per The Decoder's coverage. "We could only expose three different connectors to the model at the same time, otherwise it would forget."

The numbers are damning in retrospect. The 2023 plugin system ran on a model with a 2,000–4,000 token context window. Today, per Brockman, that window has ballooned to 52 million tokens — roughly 13,000x the capacity. The difference between "can't handle more than three tools without forgetting everything" and "can access your entire file system and hundreds of tools simultaneously."

That's not a pivot. That's a different reality.

Why This Admission Matters

OpenAI marketed plugins hard in 2023. They were framed as the next platform shift — the ChatGPT ecosystem was growing fast, developers built on it, third-party integrations like Zapier bet their go-to-market on it. The Plugin Store was supposed to be the app store moment for AI.

It lasted about a year before being quietly deprecated in favor of GPTs, which themselves faded into the background as the company pivoted hard toward agents in 2026.

Brockman's admission is notable for three reasons:

First, it confirms the timeline, not the direction, was the problem. Plugins weren't a bad idea — they were a premature one. The model couldn't hold enough context to make multi-tool workflows reliable. This is a profoundly honest take from OpenAI leadership. Most failed product postmortems blame market conditions, user behavior, or "we learned X." Brockman is saying: we shipped something the technology couldn't support, and the market correctly rejected it.

Second, it reframes the agent conversation around context, not reasoning. The narrative around agents has focused heavily on reasoning capability — can the model plan, decompose tasks, recover from errors? Brockman's plugin postmortem suggests the actual bottleneck was always context. If your model can't remember what the user asked three turns ago across four tools, it doesn't matter how good its reasoning is. The 13,000x expansion in context capacity is the real enabler of the agent era, not any single reasoning breakthrough.

Third, it's a rare moment of vulnerability from a company that usually projects inevitability. This matters for developers making platform bets. If OpenAI can admit a flagship product failed because the tech wasn't ready, it suggests the company is capable of the kind of honest retrospection that leads to better products. Compare this to the "everything is fine" messaging around products that clearly missed — and it's a refreshing signal.

From "Everything App" to "No Interface"

Here's where it gets interesting.

Brockman didn't just admit failure — he sketched the alternative. And the alternative is almost the diametric opposite of the plugin vision.

"What you ultimately want is almost no interface. You want no product," Brockman said.

He envisions ChatGPT evolving into an invisible, context-aware layer that hands off tasks autonomously. Not an app packed with features. Not a dashboard with more buttons. A persistent companion that understands your intent and executes without you needing to learn anything.

This is a notable contradiction with OpenAI's current product trajectory. The company is simultaneously:

That's a lot of product surface area for a company whose president says the goal is no interface at all.

The reconciliation is probably this: the superapp is the bridge, not the destination. You build the integrated desktop experience because the models are still jagged enough that users need a staging area — somewhere to review, confirm, and correct agent actions before trusting them fully. The "no interface" vision is the north star; the superapp is the scaffolding you need while the models earn that trust.

Brockman himself acknowledges this. Trust, he argues, is the defining product feature of the agent era — and it has to be earned, not assumed.

"Suppose you have a connector," Brockman said. "It may not send an email directly. It will say: 'I've drafted it. You need to send it.'"

What This Means for Developers Building on OpenAI

Three takeaways:

1. Bet on context capacity, not just reasoning benchmarks. Brockman's plugin story is a cautionary tale about building on platforms where the underlying model can't hold enough context to make your integration reliable. If you're building an agentic workflow today, the 500K+ token models (Claude, Gemini, GPT-5.5) are a fundamentally different development environment from the 4K-token models of 2023. Build for the context-rich future.

2. The "no interface" vision sounds revolutionary, but the intermediate steps look like traditional product work. OpenAI is shipping a superapp with desktop agents, persistent memory, and computer use. These are products with real UI surface area. The transition from "app" to "invisible layer" won't happen overnight, and there's plenty of room to build interface-layer products that help users manage their agent fleet.

3. Brockman is signaling that the company's long-term bet is on invisible infrastructure, not app ecosystems. The plugin store model — a marketplace of third-party extensions accessed through a chat interface — is dead. The replacement is context-native agents that can use any tool without explicit onboarding. If you're building an AI tool today, the winning strategy isn't "get featured in the agent app store." It's "make your tool work naturally with autonomous agents through MCP or equivalent protocols."

The Bigger Picture

Brockman's plugin admission slots into a wider pattern across the industry. Anthropic never built a plugin system — they went straight from "helpful assistant" to "coding agent with computer use." Google's agent push skips the intermediary app layer entirely. The open-source agent ecosystem has exploded precisely because developers realized they could skip the platform-wrapper step and build directly on the models.

The plugin story is a useful corrective to the hype cycle. In 2023, everyone assumed the path to agentic AI was through platform ecosystems — app stores for AI, distribution layers, curated marketplaces. Three years later, the winners are building the opposite: direct, context-rich, protocol-native agent interactions with as little interface friction as possible.

Brockman saying "the model wasn't ready" is the most honest explanation we've gotten for why the first wave of AI platforms failed. It's also the most useful framing for what comes next. The models finally are ready. The question now is whether the products can get out of their own way.