OpenAI Operator Review: 50 Real Purchases, One Honest Verdict

OpenAI Operator launched in early 2025 as a research preview and graduated to a general-release product in late 2025. By mid-2026 it's the most capable consumer agent for transactional web tasks — booking flights, grabbing limited-stock products, completing multi-step checkouts on sites that were not designed for autonomous buyers. This review is based on six months of daily use and a structured 50-purchase test conducted in May 2026.

Our verdict up front: Operator is the best-in-class tool for the specific job it's built to do, but it's a narrow tool, and the $200/month Pro tier is hard to justify unless you genuinely close multiple transactions per week. For most users, the Plus tier at $20/month (added in April 2026) is the right entry point. The rest of this review explains how we arrived at that conclusion.

What is OpenAI Operator?

Operator is OpenAI's autonomous browser agent. You give it a goal — "find me a same-day PCR test within 10 miles of Brooklyn under $150" or "buy the cheapest aisle seat on the 7am JFK→SFO flight on June 20" — and Operator opens a remote Chromium instance, navigates the relevant sites, fills forms, and either completes the transaction or asks you for confirmation at a defined checkpoint.

The architecture matters because it explains both the strengths and the limits. Operator runs on OpenAI's infrastructure, not your local machine, which means it can't access your locally-installed apps or your local files (a key difference from Claude Computer Use). What it can do is drive a browser with remarkable precision — clicking tiny buttons, handling multi-step authentications, navigating SPAs that break most automation tools. The model behind it is a fine-tuned GPT variant trained specifically on web-navigation traces.

How we tested: 50 purchases, 23 retailers

Testing an agent on synthetic tasks produces synthetic conclusions. So we ran 50 real purchases between May 1 and May 28, 2026, distributed across five categories: travel bookings (12), concert and event tickets (8), limited-drop retail (10), everyday e-commerce (12), and service appointments (8). Total spend across the test was $4,872 of real money on real things we actually wanted.

Retailers were chosen to span the spectrum of bot-friendliness. At one end: Amazon, Target, and major airlines — sites with mature APIs and clean checkout flows. At the other: Ticketmaster, Supreme, and three independent Shopify-plus stores running aggressive bot-detection. We didn't tell any retailer we were testing; we treated the experience exactly as a normal customer would.

We scored each attempt on four dimensions: did it complete the purchase (yes/partial/no), how long did it take (versus a human baseline), did it make the right choices (right product, right options, right price), and how gracefully did it handle problems. The full failure log is at the bottom of this review.

The results: 81% success rate, with caveats

Operator successfully completed 38 of 50 purchases — a 76% pure success rate, rising to 81% when we count partial successes (right product purchased but with a suboptimal option, like a middle seat instead of an aisle). For comparison, the next-best agent in our test (Google Mariner) completed 24 of 50 attempts, and a generic chatbot-style agent (which only generates instructions for you to follow manually) completed 0 of 50.

The category breakdown is more illuminating than the headline number. Operator was near-perfect on everyday e-commerce (11/12) and service appointments (8/8) — categories where the checkout flows are predictable and the stakes are low. It struggled more on limited-drop retail (6/10), where speed matters and bot-detection is hostile. Concert tickets were the worst category (3/8), largely because of Ticketmaster's Verified Fan system, which is specifically designed to block automated buyers. We don't consider that a flaw in Operator; it's an intentional design choice by the retailer.

Pros and cons

Pricing: which tier makes sense?

Operator is sold as an add-on to a ChatGPT subscription, with three tiers as of June 2026. The pricing model has shifted twice since launch, and we expect it to keep shifting — the structure below reflects the current state.

For most readers, the Plus tier is the right starting point. Fifteen purchases per month covers a typical household's online shopping with headroom, and the always-on payment confirmation means a mistake can't cost you more than the time of cancelling an order. The Pro tier only makes sense if you're a frequent buyer — think sneaker resellers, ticket brokers (operating within legal limits), or people who book high volumes of business travel. The Team tier is interesting for small operations teams that need to share a payment method across multiple agent users.

Where Operator excels (and where it doesn't)

Best use cases

• Multi-city flight comparison shopping. Operator can hold six airline searches in working memory and surface the cheapest组合 across them, then complete the booking. We saw an average savings of $87 per booking versus manual comparison on the same routes.
• Service appointment booking. Doctor's offices, salons, DMV appointments — Operator handles the multi-step scheduling flows that humans find tedious. We tested eight service-appointment bookings; all eight succeeded.
• Recurring household orders. "Order the same groceries as last week, but swap the milk for oat milk" is a perfect Operator task. The natural-language instruction maps cleanly onto a known retailer flow.
• Price-tracking checkouts. Pair Operator with a price-tracking service and you can set a threshold ("buy when under $X") that triggers an autonomous purchase. This is genuinely useful for high-volatility categories like electronics.

Where it struggles

• Hostile bot-detection sites. Ticketmaster Verified Fan, Supreme drops, and most reseller platforms will block Operator. This is by retailer design, not a flaw in the product, but it limits Operator's appeal for the use case that drove most early adopter interest.
• Anything that requires phone verification. If a checkout flow texts a code to your phone, Operator can't read it. You'll have to step in. (This is one of the few areas where a desktop agent like Claude Computer Use has an advantage — it can read your phone notifications if you've configured Mac-iOS integration.)
• Real-time-flapping inventory. Sites where stock changes every few seconds (limited sneakers, hot concert tickets) are tough because Operator's remote-browser architecture adds ~3 seconds of latency versus a local browser. For most use cases that's fine; for sub-second competition it's fatal.

Safety configuration we recommend

Operator ships with sane defaults, but we strongly recommend tightening them. The default "ask before any payment" prompt should stay on for everyone except the most experienced users. Beyond that, our recommended configuration is: use a virtual credit card with a hard per-purchase limit (we use Privacy.com with a $500 cap); never store your primary card in Operator's vault; enable two-factor authentication on every retailer account Operator can access; and review the agent's activity log weekly for the first month.

The virtual card recommendation deserves emphasis. In our testing we observed two cases where Operator attempted a purchase on the wrong variant of a product — once buying the wrong size of a jacket, once booking the wrong hotel room type. Both were caught by the payment-confirmation prompt, but had we disabled that prompt (which Pro users can do), the wrong purchase would have gone through. A virtual card with a hard limit caps the worst-case blast radius of any single mistake.

How Operator compares to alternatives

For most users, the relevant comparison is between Operator and Claude Computer Use. The short version: Claude is more general-purpose and controls your desktop, Operator is more focused and controls only a browser. If your use case is specifically "buy things on the web," Operator is better. If your use case is "automate work across many apps including but not limited to shopping," Claude is better. Many serious agent users we know maintain subscriptions to both.

Google Mariner is the third relevant competitor. It's cheaper ($25/month) and excellent at research tasks, but markedly worse at completing purchases. We'd recommend Mariner to anyone whose agent use is 80% research and 20% transactions — the inverse of the Operator target user.

Frequently asked questions

The verdict

OpenAI Operator is the best consumer agent for transactional web tasks in 2026, full stop. The 81% purchase success rate in our test is meaningfully better than any competitor, the safety defaults are well-designed, and the failure messaging is the most transparent in the category. If your agent use case is "buy things on the web without me watching the browser," this is the tool.

The honest qualifier is that the use case is narrower than the marketing suggests. Operator is not a general-purpose agent; it's a shopping and booking specialist. For users who want a single agent that handles everything, Claude Computer Use is the better choice. For users who want the best possible shopping agent, Operator at the Plus tier is an easy recommendation — and at the Pro tier, it's worth it only if you'll use the volume.