OpenAI can’t stop accidentally leaking GPT-5.4. On Monday evening, multiple developers caught a cybersecurity block inside Codex that referenced a model called gpt-5.4-ab-arm1-1020-1p-codexswic-ev3.

This wasn’t a one-off fluke. Two separate pull requests in OpenAI’s public Codex GitHub repo referenced GPT-5.4 by name—one even included a slash command to "toggle Fast mode for GPT-5.4." Both were scrubbed within hours via force-pushes, but not before the community screenshotted everything.

The Leaked GPT-5.4 Model Selector

Minutes after the GitHub scrub, OpenAI's own Codex employee (Tibo) briefly posted—then deleted—a screenshot showing GPT-5.4 in the model selector. When asked about it, he jokingly told the community, "You saw nothing." We are now at five major GPT-5 variants in seven months. At this rate, GPT-5.9 ships before your Q2 OKRs.

The leaked specs suggest OpenAI is moving to end the "Goldfish Memory" era. GPT-5.4 reportedly features a 2,000,000 token context window, specifically designed to match Google’s Gemini 3.1 Pro.

But the real kicker is "Stateful AI." Unlike current models that start fresh every time, Stateful AI retains the state of your workflow, development environment, and tool calls across sessions. It doesn't just remember your text; it remembers your entire project environment like a real-life colleague sitting next to you.

Lost in the leak madness was the official release of GPT-5.3 Instant. OpenAI finally addressed the biggest user complaint of 2026: the "cringe." The new model significantly reduces "preachy" disclaimers, moralizing preambles, and unnecessary refusals. It’s faster, leaner, and—thankfully—much less annoying to talk to.

GPT-5.3 Instant is a high-velocity reasoning model designed to eliminate the conversational friction of its predecessors by removing "preachy" moralizing and unnecessary disclaimers. It delivers near-instantaneous response times with a refined focus on direct, instruction-following accuracy for technical and creative workflows.

The Death of the "Chatbot" Era Insiders analyze that skipping straight to 5.4 is OpenAI’s attempt to pivot from "Chat Assistant" to "Digital Employee." With a 2M context window and stateful memory, they aren't building a tool you talk to—they’re building an autonomous agent that works for you. This comes as Anthropic’s Claude 4.6 continues to dominate the "Vibe Check" on social media, forcing Sam Altman to speed up the release cycle to a near-weekly cadence.

The developers. The new Codex App for Windows is OpenAI’s first native desktop application, delivering low-latency agent threads that bypass the browser to interact directly with your local file system and IDE. By integrating with the Windows kernel, it enables "Native Context," allowing multiple autonomous agents to execute terminal commands and monitor your workspace in real-time. This effectively transforms your local machine into a high-intensity, collaborative AI workstation for professional development.

⚡ The Vibe Check: An arena watcher noted that GPT-5.4 has already been spotted in the wild for some Pro users, claiming it makes current models "feel like calculators." Translation: The capability gap is about to get very wide, very fast.

Today was the day the version number war went nuclear. While everyone was testing the "less cringe" GPT-5.3, OpenAI accidentally showed its cards for the next leap. 2 million tokens and persistent memory isn't just an upgrade; it's a reset. If you’re still prompting like it’s 2024, you’re already behind. The era of the autonomous "Digital Employee" starts now.

Stay building. 🤖