← Back to all agents
Browsing Commercial
ChatGPT Operator
OpenAI's browser agent that can click and fill forms for you.
Our take
Impressive on routine web tasks, but expensive and still early.
What it is
Operator is OpenAI's browser-using agent. It opens a remote browser, sees the screen, and can navigate, click, type, and fill forms to complete a task on your behalf. Useful for repetitive web tasks like booking, ordering, or filling standardized forms, though it can still get stuck on unusual UIs.
Best for
- Fill in repetitive online forms
- Book simple appointments (restaurants, services)
- Place routine online orders
Pros
- Can use the web like a human — no API integrations needed
- Built on a strong vision model that handles most UIs
Cons
- Slow (it thinks in steps, not seconds)
- Expensive — bundled with $200/mo ChatGPT Pro
- Still struggles with CAPTCHAs and unusual layouts
Quick start
- Subscribe to ChatGPT Pro ($200/mo) — Operator is bundled in. Plus / Team access is rolling out.
- Open ChatGPT → use case → "Run tasks with Operator" → describe what you need (e.g. "Find the cheapest direct flight from SFO to Tokyo between Nov 10–17").
- Operator opens its own virtual browser, navigates with mouse + keyboard, asks you to confirm sensitive steps (payments, login).
- July 2026 update: rolled into the new "ChatGPT Agent" which combines Operator + Deep Research + ChatGPT into one tool that picks tools on its own.
- On benchmarks: Humanity's Last Exam 41.6% / 44.4% with parallel votes; BrowseComp 68.9% (SOTA).
Sample input / output
Input
Task: "Browse the NFL schedule for the 2025 regular season and plan a route to see all 30 MLB teams ballparks, including hotels and a final CSV + map." Output
Operator + Deep Research workflow:
- ~10 min runtime in its virtual computer
- Output: Excel itinerary + interactive map visualization (slides-ready)
- Visible screen-recording of every browser action
- Tools chosen dynamically: visual browser for booking sites, text browser for long-form schedules, terminal for CSV cleanup Benchmarks
Humanity's Last Exam (single pass) 41.6% (was 26.6% for Deep Research alone) — OpenAI July 2026
BrowseComp 68.9% (SOTA; +17.4 vs Deep Research) — OpenAI July 2026
FrontierMath 27.4% accuracy — OpenAI July 2026
Investment-banking analyst tasks Matches / beats junior IB associates (1–3 yrs experience) — OpenAI demo, July 2026
Underlying model (May 2025) CUA, built on o3 — 至顶网
Pricing
Included with ChatGPT Pro ($200/mo)
Underlying models
Should you pick this?
Pick it if You already pay for ChatGPT Pro and want hands-off browser tasks — Operator (now ChatGPT Agent) gives you the best-in-class UX and is included in the $200/mo plan.
Skip it if You want self-hosting, fine-grained control, or a programmatic API — Browser Use, Firecrawl, and Playwright give you building blocks at a fraction of the cost.