OpenAI Operator Review 2026: Features, Pricing & Verdict

Operator is included in ChatGPT Pro, and Pro now starts at $100 a month, not $200. That is most of the buying decision. If you already pay for Pro, Operator is capability you have already bought — use it. If you are on Plus at $20, do not upgrade for Operator alone: Manus Pro at $20 is better at autonomous research, and Plus plus Manus costs $40 against Pro 100’s $100. Upgrade only for a recurring browser task no API can reach — the same form across hundreds of sites, a booking flow, weekly competitor price checks — and only after you have measured that it runs often enough to justify $80 a month of delta. The vision-action quality is the best sold to consumers. The subscription is also the only way to get it.

What it costs, and what the ladder buys

Operator requires a Pro tier. Since DevDay on 29 September 2026 Pro is three plans rather than one:

Plan	Monthly	Included usage	Operator
Free	$0	baseline	No
Go	$8	same entitlement as Free	No
Plus	$20	1x	Waitlist only
Pro 100	$100	5x Plus	Yes
Pro 200	$200	10x Plus from 30 Oct 2026	Yes
Pro 500	$500	25x Plus; only tier with Ultrafast	Yes

Prices verified 23 October 2026 against OpenAI’s pricing docs, the only reachable OpenAI pricing surface — chatgpt.com/pricing and both help.openai.com plan articles return 403. The allowance multiples are reported rather than published; OpenAI’s live page says only that “Pro plans currently have no five-hour limit.”

The arithmetic that matters for Operator is what happens on 30 October 2026, when Pro 200’s included usage falls from 20x Plus to 10x at an unchanged $200. After that date every rung costs the same $20 per 1x of Plus allowance, so Pro 200 buys no volume discount over Pro 100 — only headroom. For browser automation that is the wrong direction, because Operator is the most allowance-hungry thing in ChatGPT: every step of a task is a fresh screenshot, a reasoning pass and an action, and a cross-site task runs 25–40 minutes. Buy Pro 100 and measure before you buy Pro 200. Note too that speed modes multiply consumption against the same allowance — Fast bills 2.5x, Ultrafast 8x — so running Operator in Fast mode on Pro 100 spends a 5x allowance at 2.5x the rate.

What Operator would cost by the token

OpenAI does not sell Operator by the token, which makes it hard to know whether $100 is cheap or dear. GPT-6 Astra, the model underneath computer use, does have a rate card: $10 per million input tokens, $1 cached, $50 output, with a threshold at 272,000 input tokens above which input and cache bill at 2x and output at 1.5x. Browser-agent loops accumulate context by design — each step resends the running transcript — so the cost grows with the square of the step count.

Model an 80-step cross-site task whose context grows about 6,000 tokens per step, with 90% of each request served from cache:

• Context crosses 272,000 tokens at step 46, so the second half of the task bills at the doubled rate.
• Steps 1–45 bill about 6.2M input tokens: $11.80.
• Steps 46–80 bill about 13.2M input tokens at 2x: $50.27.
• Output across 80 steps is trivial by comparison: about $1.46.
• Total: roughly $64 for one long task.

Two such tasks a month would consume more than Pro 100’s price in tokens alone. That is the honest case for the subscription — and the reason OpenAI prices Operator as an entitlement rather than a meter. Run the same arithmetic on your own step counts with the API cost calculator.

What Operator is

Operator is OpenAI’s computer-using agent. It is a loop: a vision model captures the Chrome window, a reasoning model picks the next action, an executor emits mouse coordinates and keystrokes, repeat until done or interrupted. Because it reads pixels rather than the DOM, it can navigate any site you can, including sites that block conventional automation.

In 2026 OpenAI folded three capabilities into one ChatGPT product: Operator’s browser control, Deep Research’s synthesis and ChatGPT’s conversation. You do not invoke “Operator mode”; you ask for the outcome and the orchestration is hidden. GPT-6 Astra, which shipped on 3 September 2026, is the model that drives it — 72.6% on OSWorld 2.0 against GPT-5.6 Sol’s 65.7%, 92.7% on ScreenSpot-Pro against 76.9%, and roughly 47% less wall-clock time per task in OpenAI’s own latency simulations.

The successor line is dots, launched 29 September 2026: always-on Astra agents with their own cloud computer and browser. Their proactive research is read-only by design, and personal Pro access excludes the EEA, Switzerland and the UK at launch, while Business Premium reportedly carries them across all supported regions. If you are an individual buyer in Europe, the capability Operator is migrating into is not currently sold to you.

Nobody sells a computer-use API any more

The review used to say “there is no Operator API” as a quirk. It is now a market fact with dates. OpenAI shipped a dedicated computer-use-preview model, deprecated it on 22 April 2026 and shut it down on 23 July 2026, naming the general-purpose gpt-5.6-terra as its replacement rather than another computer-use model. Google’s gemini-2.5-computer-use-preview-10-2025 shut down five days later, on 28 July 2026, replaced by gemini-3.8-flash. Both dates are in the model retirement calendar.

So within one week both vendors withdrew the API layer for computer use and kept the capability inside their consumer products. The practical consequence for a buyer: if you want to build on browser control rather than use it, your remaining options are Anthropic’s computer-use tool, which is still sold as an API tool, or a third-party service such as Browse AI. You cannot buy OpenAI’s.

What Operator does well

Browser tasks no API can reach. Booking through travel sites with no public API. Filling forms across 50 prospect sites. Weekly competitor price checks. No credentials, no scraping cooperation.

Cross-site workflows. “Compare this part across three retailers, then book calls with the top two vendors” runs end to end and reports back once.

Survives layout changes. When a button moves, the vision model reads the new screen. Selector-based scrapers break; Operator usually adapts.

Visible execution. You watch it work and can intervene. Genuinely safer than blind agent execution, and the main reason trust in it is higher than its reliability alone would justify.

Where Operator falls short

Still a $100 floor. Halved, but not cheap, and Plus access remains theoretical.

No API, by vendor decision. See above — this is now permanent rather than pending.

Slow. Screenshot-plus-reason-plus-act is compute-heavy. Tasks Manus finishes in 10 minutes take Operator 25–40.

Blocked by the usual walls. CAPTCHA, login gates and aggressive bot detection stop it like they stop a human, and its retry behaviour is wasteful when they do.

Detectable. Its browser fingerprint is identifiable, so sites that want to exclude AI agents can.

No memory across runs. Every task starts cold unless you paste the context in again.

Not a team product. Built for one person running one task. Production team workflows need Lindy or an agent SDK.

Who should buy which

You are	Buy	Why
Already on Pro for other reasons	Nothing	Operator is included; the cost is sunk
Doing autonomous research	Plus + Manus Pro, $40	Denser, better-cited output at 40% of Pro 100
Running a recurring browser task	Pro 100, $100	The only way to buy Operator; measure before Pro 200
Hitting Pro 100’s 5x allowance weekly	Pro 200, $200	Headroom only — no volume discount after 30 Oct
Needing Ultrafast	Pro 500, $500	The one tier that adds a capability
Building browser control into a product	Claude’s computer-use tool	Anthropic still sells it as an API
In the EEA, Switzerland or the UK	Pro 100 for Operator, not dots	Personal Pro excludes dots in those regions

Full ranked picks at best AI agents in 2026, and current plan prices for every tool in the pricing tracker.

The verdict

Operator is the most polished computer-use agent an individual can buy, and the price wall that defined this review has halved — $100 rather than $200 is the difference between “obviously not” and “worth doing the arithmetic”. Do the arithmetic anyway. Pro 100 pays for itself only against a task you run repeatedly and cannot automate another way. For everyone else, Plus plus Manus at $40 covers most of the ground, and the token model above shows why OpenAI would rather sell you an entitlement than a meter.

One thing to watch rather than buy: dots is where this capability is going, and the EEA, Switzerland and the UK are currently outside the personal-plan version of it.

Common questions

Is Operator still $200 a month?

No. The $200 figure dates from when ChatGPT Pro was a single plan. Since 29 September 2026 Pro sells at $100, $200 and $500, and Operator is available on all three, so the entry price for computer-use work is $100.

Can I use Operator on ChatGPT Plus?

Not in practice. Plus subscribers can join a waitlist, and even when granted, the usage limits make repeated browser-automation work impractical. The unified ChatGPT Agent flow is more accessible on Plus than Operator-specific browser control is.

Should I buy Pro 200 instead of Pro 100 for heavy browser automation?

Only for headroom. From 30 October 2026 Pro 200’s included usage halves from 20x Plus to 10x at an unchanged price, so every Pro rung costs the same $20 per 1x of allowance. Pro 200 gives you twice Pro 100’s capacity for twice the money and adds no capability; Pro 500 is the only tier that adds one, namely Ultrafast.

Is there an Operator or computer-use API?

Not from OpenAI. Its computer-use-preview model was shut down on 23 July 2026 and replaced by a general-purpose model, and Google retired its Gemini computer-use preview on 28 July 2026. Anthropic still ships computer use as an API tool, which makes it the default choice if you are building rather than buying.

Why is Operator slower than other agents?

Because every step costs a screenshot, a reasoning pass and an action, and the context carries forward. That is also why cost grows with the square of the step count — a long task spends its second half above GPT-6 Astra’s 272,000-token threshold, where input bills at double.

Can I get dots in Europe?

Not on a personal Pro plan at launch: dots excludes the EEA, Switzerland and the UK for Pro users. Business Premium reportedly carries them across all supported regions.

What changed

• 23 Oct 2026: page restructured answer-first; Pro tiers re-verified at $100/$200/$500 and the “no API” claim dated — OpenAI’s computer-use-preview shut down 23 July 2026 and Google’s Gemini computer-use preview on 28 July 2026, so no vendor but Anthropic sells the capability as an API.
• 1 Oct 2026: Pro 500 launched and Pro 200’s allowance halves to 10x on 30 October, flattening the ladder to $20 per 1x — see analysis.
• 1 Oct 2026: dots shipped as Operator’s successor line, with personal Pro access excluded in the EEA, Switzerland and the UK — see analysis.
• 29 Sep 2026: ChatGPT Pro split into three tiers, moving Operator’s entry price from $200 to $100.
• 4 Sep 2026: GPT-6 Astra became the model under computer use, at $10/$50 per million tokens with a 272,000-token threshold that doubles input — see analysis.
• 23 Jul 2026: OpenAI shut down computer-use-preview, ending any prospect of an Operator API.