Three days ago OpenAI posted "5.4 sooner than you think" on X and walked away. Today they delivered. GPT-5.4 is live right now in ChatGPT, Codex, and the API. And this isn't a minor version bump. This is OpenAI bringing everything they've built over the last six months into one unified model. Reasoning. Coding. Computer use. Professional work. All in one place.

GPT-5.4 Standard replaces GPT-5.2 as the default for Plus, Team, and Pro users. Free for everyone today.

GPT-5.4 Thinking shows you its full reasoning plan before answering so you can redirect it mid-response instead of starting over.

GPT-5.4 Pro is the maximum performance tier built for Enterprise teams who need the absolute best.

Computer Use: GPT-5.4 scored 75% on OSWorld-Verified, beating human testers who average 72.4%. Claude Opus 4.6 sits at 72.7%. GPT-5.4 wins.

Real Work Tasks: 83% win or tie rate across 44 professions including law, finance, medicine, and engineering. That's not a benchmark. That's your job.

Scientific Reasoning: 94.4% on GPQA Diamond with no tools allowed. Expert level science, solved without looking anything up.

Advanced Math: 50% on FrontierMath Tier 1-3 and 38% on Tier 4. Six months ago every AI lab on earth scored under 5% on this same test.

Software Engineering: 57.7% on SWE-Bench Pro. The highest OpenAI has ever scored on coding tasks.

The One They Lost: BrowseComp, agentic web browsing. Gemini 3.1 Pro still leads at 85.9% vs GPT-5.4's 82.7%. Google holds one crown. For now.

GPT-5.4 is the first general-purpose OpenAI model with native computer use built in. It can read your screen, issue keyboard commands, control your mouse, and navigate desktop apps autonomously. Not in a demo. Not in a sandbox. Live today in the API and Codex. You point it at your computer, tell it what to get done, and walk away. Think about what that actually means for your day. Every repetitive task that lives across multiple apps, every copy paste workflow, every "open this, find that, move it here" process you do on autopilot. GPT-5.4 can handle all of it while you focus on the work that actually requires you. This is the feature that turns AI from a chat window into an actual employee.

This is the one that broke the internet today. OpenAI demoed GPT-5.4 building a fully playable theme park simulation game from a single prompt. Not a concept. Not a wireframe. A working game.

What it built on its own: Tile-based path placement, ride construction, guest pathfinding, queueing systems, ride cycles, and a full park economy tracking money, happiness, cleanliness, and guest ratings. Every visual asset generated by the model itself.

Then it playtested it. GPT-5.4 ran the game autonomously, expanded the park across multiple rounds, verified navigation worked, checked that guests and queues updated correctly, and flagged its own bugs before a human ever touched it.

One prompt. One model. A complete game.

If you're a developer reading this right now, you already know what this means.

The API version supports up to 1 million tokens of context. The entire Lord of the Rings trilogy is about 500,000 words. GPT-5.4 can hold two of those in memory simultaneously and still have room left over. For enterprise teams working with massive codebases, lengthy contracts, or deep research libraries this changes what's possible in a single session. No more chunking documents. No more losing context halfway through a complex project. No more "as I mentioned earlier" moments where the model forgot what you said twenty minutes ago. You feed it everything at once and it holds it all. That's not an upgrade. That's a different category of tool entirely.

A week ago OpenAI was dealing with public backlash, losing users to Claude, and watching Anthropic dominate every headline. Today they shipped the most capable AI model ever built by any company. The AI war doesn't pause for controversy. It doesn't pause for politics. It doesn't pause for anything. GPT-5.4 is today's headline. Something else will be next week's. The only question is whether you're keeping up.

Stay building. 🤖