I genuinely don't know how to feel about this one. OpenAI just released GPT-5.4. Half my feed is losing their minds over the benchmarks, and the other half is already complaining about API latency. But looking past the noise, this release signals a clear shift. We are moving away from models that just chat, and moving toward models that actually do things.
GPT-5.4 is pitched specifically for "professional work." It pulls in the coding chops from GPT-5.3-Codex, improves general reasoning, and adds a few new mechanics that make agentic workflows much more viable.
Here is a breakdown of what actually matters in this update.
Native computer use is here
The biggest news is that GPT-5.4 has native computer-use capabilities. It can look at screenshots and issue mouse and keyboard commands to navigate a desktop environment.
On the OSWorld benchmark, which tests a model's ability to navigate a computer, GPT-5.4 scored 75%. For context, GPT-5.2 was at 47.3%, and human performance sits around 72.4%. It is technically beating average human baselines at point-and-click tasks.
There is something unsettling about agents churning through spreadsheets and web portals while nobody is watching. But from a productivity standpoint, the value is obvious. Developers can configure the model's safety behavior and set confirmation policies so it does not accidentally delete production databases.
Tool search fixes the context window problem
If you have tried building agents with lots of tools, you know the pain of dumping every single tool definition into the system prompt. It eats up tokens, drives up costs, and slows down responses.
GPT-5.4 introduces "tool search" in the API. Instead of receiving all tool definitions upfront, the model gets a lightweight list. When it needs to use a tool, it looks up the definition and pulls it into the context dynamically.
OpenAI claims this reduced token usage by 47% on the MCP Atlas benchmark. This is a nerdy backend detail, but it changes everything for developers. You can now build agents with thousands of available tools without bankrupting your API budget.
Steerable thinking in ChatGPT
You know that feeling when a model starts generating a massive response, and halfway through you realize it misunderstood the prompt? You usually have to hit stop and start over.
With GPT-5.4 Thinking in ChatGPT, the model now provides an upfront plan of its thought process. If you see it going off the rails, you can adjust its course mid-response. You just correct it while it is working, and it incorporates your feedback into the final output. This alone will save me hours of wasted generation time.
The model is also better at maintaining context during long, complex tasks. It does not lose the plot as easily when you ask it to synthesize information from multiple web searches.
Pricing and availability
GPT-5.4 is rolling out today. If you are a ChatGPT Plus, Team, or Pro user, you have access to GPT-5.4 Thinking right now. The old GPT-5.2 Thinking will stick around in the legacy menu until June.
In the API, the base GPT-5.4 model costs $2.50 per million input tokens and $15 per million output tokens. That is a price bump from GPT-5.2. But OpenAI argues that the model uses fewer tokens to solve problems, so your total bill might actually go down. We will see how that plays out in reality.
There is also a GPT-5.4 Pro version for heavy-duty tasks, which runs $30 for input and $180 for output per million tokens.
Official Links
- Project Page / Demo: https://openai.com/index/introducing-gpt-5-4/
Time to build
This update feels less like a magic trick and more like a tool upgrade. The AI is getting better at using the same software we do.
If you are building agents or just tired of copying and pasting data between tabs, this model is worth testing. You can try it out in ChatGPT now, or start plugging it into your API workflows.