Key takeaways

  1. Agent mode is on for all Plus ($20/mo) and Team accounts as of 2026-05-27. Enterprise rollout follows next week.
  2. Inside agent mode: browser, Python sandbox, file system, scheduled runs. Outside: nothing — no email send, no purchases, no API calls without per-action approval.
  3. Practical use cases that work today: research compilations, repeatable spreadsheet workflows, monitoring tasks with notifications.
  4. Practical use cases that don't work: real-time anything, anything requiring external authentication, complex multi-tab workflows.
  5. Pricing didn't change. Agent runs count against your existing usage limits.

OpenAI flipped on GPT Agent Mode for all Plus and Team subscribers this morning, finishing a rollout that had been in limited beta since March. The capability is exactly what it sounds like — give GPT a multi-step goal, and it figures out which tools to use, in what order, and reports back when done — and it’s been the centerpiece of OpenAI’s “agentic computing” narrative for the last year.

The interesting story isn’t the launch itself, which was telegraphed. It’s how OpenAI scoped the release. After several rounds of safety incidents in the beta, the production version is dramatically more conservative than the demos suggested.

What you can do

Inside the sandbox, agent mode has access to:

  • A full browser. It can search, navigate, fill forms, scroll, click. It cannot log in to anything that needs your real credentials (separate flow required).
  • A Python environment. Code execution, file I/O, data analysis. Persistent across runs in the same chat.
  • A scratch file system. Documents, spreadsheets, images. Files can be downloaded but not auto-emailed.
  • Scheduled runs. Set an agent to run every weekday morning, or weekly, or on-demand. Results land in a chat.

Genuinely useful tasks the rollout handles well:

  • “Pull the latest releases from these 12 GitHub repos and summarize what changed.”
  • “Find me five remote senior product manager openings in fintech and put them in a spreadsheet with company, salary range, and a 3-bullet summary of each.”
  • “Every Monday at 9am, check these five AI news sites and email me a one-page digest.” (Email currently sends to your ChatGPT inbox, not your real email — see caveats.)

What you can’t do

The conservative scoping comes through in what’s blocked or gated:

  • No external writes by default. Booking a flight, sending an email outside the ChatGPT environment, posting to a social account — all require per-action user approval.
  • No long-running connections. Sessions max out at 30 minutes; agents that need persistent state have to checkpoint.
  • No real-time anything. The sandbox doesn’t have a stable IP for things like webhook listening or streaming data.
  • Authentication is awkward. You can’t pass it credentials directly. OAuth-style integrations are gated to a curated allowlist (Gmail, Calendar, GitHub, Notion, Linear, and a few others at launch).

These are the right defaults — the beta had several incidents where agents made unintended purchases or emails — but they constrain the use cases significantly. If you’ve been told an agent will book your travel autonomously, that’s still mostly aspirational. It will find the flights and put them in a comparison sheet. The actual booking still needs you.

It will find the flights and put them in a comparison sheet. The actual booking still needs you.

Pricing and limits

No new pricing tier. Agent mode counts against your existing usage budget. OpenAI hasn’t published exact rate cards but a Plus user can expect to run roughly 25-50 agent tasks per day before hitting limits, depending on complexity. Team users get higher caps.

The compute is non-trivial — agent runs often involve dozens of LLM calls plus browser and Python overhead — so heavy daily use of agents is likely the trigger that pushes people from Plus to Team or Enterprise.

How it stacks up

The two natural comparisons:

  • Claude Computer Use (Anthropic, launched late 2025) is more capable inside a desktop sandbox but harder to schedule and lacks the curated integrations.
  • Gemini Workspace Agent (Google, in beta) is tightly integrated with Google Workspace but limited to Workspace-shaped tasks.

GPT Agent Mode lands between them: more polished than Claude’s, broader than Gemini’s, but more restricted than either on raw capability. For most users this trade is correct. For developers building production agents, the API (still in limited preview) will matter more than the consumer rollout.

What to do today

If you’re a Plus or Team subscriber, the practical move is:

  1. Try a research-style task first. Something well-bounded, with a clear deliverable (a spreadsheet, a doc, a list).
  2. Don’t try to automate anything that touches money, identity, or external sending. Those flows aren’t ready.
  3. Set up one scheduled run — a weekly digest or report. The scheduling layer is the thing most people will derive ongoing value from.
  4. Watch your usage. Agents burn through quota faster than chat.

The capability is real, even if the early use cases are narrower than the keynote suggested. The interesting six months ahead will be watching what people actually build with this — and which of the conservative defaults OpenAI starts to loosen.

About Aditya Marin Gasga

Founding Editor

Aditya covers the whole AI surface area for Signal — frontier models, agent infrastructure, the economics of inference, and the policy decisions that quietly shape what everyone else can build. He writes for operators who need a calibrated view of what's actually shipping versus what's keynote theatre.

  • Founder of Signal; sets the publication's editorial line
  • A decade across product, growth, and AI tooling at venture-backed startups
  • Reads the model release notes, the system cards, and the benchmark papers — and tells you which ones matter
More from Aditya Marin Gasga →