Check the calendar, research the news, and create a briefing document — all three done with a single prompt. The ChatGPT Agent that OpenAI first unveiled in July 2025 became a completely different product in March 2026 when it was equipped with GPT-5.4. Operator, which had been running as a separate service, is set to be discontinued, and Deep Research’s analytical power has been absorbed into the agent. It’s now truly an "all-in-one" AI.
What Is This?
Let me quickly walk through the timeline. In January 2025, OpenAI launched Operator — a service where AI directly controls a browser. Around the same time, there was Deep Research — a research-specialized feature that reads dozens of sources and creates comprehensive reports. The problem was that these two operated completely separately.
Operator was great at clicking and scrolling websites but clumsy at reading and analyzing long documents. Deep Research was a master analyst but couldn’t even access sites that required login. In July 2025, OpenAI merged both teams to create ChatGPT Agent.
Then on March 5, 2026, GPT-5.4 launched and ChatGPT Agent entered an entirely new phase. GPT-5.4 is OpenAI’s most powerful frontier model. It’s the first to unify reasoning, coding, and agent workflows in a single model, and the first to ship simultaneously to ChatGPT, API, and Codex.
The key change is native Computer Use. GPT-5.4 can directly interpret screens and control mouse and keyboard to automate complex workflows. Text browser, visual browser, terminal, API integrations — all tools share a single state and switch seamlessly.
The inside story shared by OpenAI researchers on the Sequoia Capital podcast is impressive. The team that built the agent was surprisingly small. 3–4 Deep Research researchers, 6–8 Operator researchers, and an applied engineering team. This small group used reinforcement learning (RL) to train across thousands of virtual machines on diverse tasks. The key was that "they didn’t prescribe tool-use patterns — they let the model find the optimal strategy on its own."
Operator Being Discontinued
With the virtual browser now built into ChatGPT Agent, the separately operated operator.chatgpt.com is set to be discontinued within weeks. Existing Operator users can simply switch to ChatGPT’s agent mode with no additional steps needed.
What Changes?
The previous ChatGPT Agent (July 2025) could already manipulate browsers. But what changed with GPT-5.4 is that it went from "can do it" to "does it well."
| Before (GPT-5.2 based) | Now (GPT-5.4 based) | |
|---|---|---|
| Reasoning Model | GPT-5.2 Thinking + o3 | GPT-5.4 Thinking single model |
| Desktop Control | Primarily web browser | Native computer use (mouse + keyboard) |
| Expert-Level Tasks | 70.9% match across 44 professions | 83.0% match across 44 professions (GDPval) |
| Coding | Basic code generation | GPT-5.3-Codex-level coding + frontend polish |
| Context | Limited | 272K default, up to 1M tokens |
| Spreadsheets/PPT | Basic generation | Directly creates and edits editable files |
| Thought Process | Black box | Shows thinking plan upfront, user can adjust |
The OSWorld-Verified benchmark is symbolic. It measures AI’s ability to perform tasks in a real desktop environment, and GPT-5.4 scored 75% — surpassing the average human score of 72.4%. It’s the first benchmark proof that "AI can use a computer better than humans."
Sequoia Capital’s analysis nails it — they called it "1+1=3." Combining Operator’s visual web manipulation with Deep Research’s text analysis and synthesis lets it accomplish things neither could do alone. For example, if you say "analyze 3 competitors and make a slide deck" — it browses websites and collects data (Operator capability), synthesizes the collected information (Deep Research capability), and generates an editable presentation file (new GPT-5.4 capability) — the entire process runs as one.
Things to Know
While GPT-5.4 is a major improvement, agent tasks still take 5–30 minutes. It can still stumble on simple UI elements like datepickers, and adapting to different website layouts remains a challenge. Always manually verify high-stakes actions like payments or sending emails.
The Essentials: How to Get Started
- Check your plan
GPT-5.4 Thinking is available to ChatGPT Plus ($20/mo), Team, and Pro ($200/mo) users. Plus gets 40 agent runs/month, Pro gets 400. GPT-5.4 Pro (highest performance) is exclusive to Pro and Enterprise. - Enter agent mode
Select "agent mode" from the tool dropdown at the bottom of the ChatGPT chat, or type/agent. You can switch mid-conversation at any time. - Connect app connectors
Link Google Calendar, Gmail, Google Drive, GitHub, etc. to enable personalized tasks like "Check my calendar and brief me on next week’s meetings." - Assign your first task
Research + organization combos have the highest success rate. Try things like "Find 5 AI news stories from this week and create a summary table" or "Compare pricing for competitors A/B/C and organize it in a spreadsheet." - Automate recurring tasks
Click the clock icon on a completed task to schedule daily/weekly/monthly repeats. Try automating "Every Monday morning, competitor news briefing."
