Shootout #3Results in
Find the best price on a Ninja AF101 air fryer
Four agents, same prompt, scored Nov 3, 2026. Ranked by points, then items met, then fewest factual errors, then time. Every scored run cost $0 extra and needed 0 interventions. No agent added to a cart, signed in, or contacted anyone. Lindy did not run.
Results
ChatGPT and Genspark each met 8 of 8 and take 2 points. Each has 1 factual error, so ChatGPT is first on time (1.3 min). Manus and Grok each met 7 of 8 and take 1 point. Manus has 0 factual errors and Grok has 3, so Manus ranks ahead.
| Rank | Agent | Result | Pts | Items | Time | Interventions | Cost | Plan |
|---|---|---|---|---|---|---|---|---|
| 1 | ChatGPT agentWork (GPT-6.1 Sol, Light) | Pass | 2 | 8/8 · 1 factual error | 1.3 min | 0 | $0 | ChatGPT Plus |
| 2 | GensparkStandard (free) | Pass | 2 | 8/8 · 1 factual error | ~5 min | 0 | $0 | Free |
| 3 | ManusFree | Partial | 1 | 7/8 · 0 factual errors | 1.9 min | 0 | $0 | Free |
| 4 | GrokHeavy | Partial | 1 | 7/8 · 3 factual errors | ~1.2 min | 0 | $0 | SuperGrok |
Did not run
A missing run is not a score. It is not a fail, and it is not in the pass-rate count.
Lindy
Did not run: free trial expired, needs paid plan. Free trial expired; requires $29.99/mo plan. Skipped per no-new-subscriptions rule. Recorded as 0 runs, not a scored run.
Notes
ChatGPT agent
Fast, clean table and a sound pick: Best Buy ($119.99). Its $119.00 Walmart marketplace offer (Athena's Amazing Finds) could not be found on the page, so prices are flagged. The Walmart seller was unverified, so seller and return window is flagged. Run
- Cited a $119.00 Walmart price from Athena's Amazing Finds. That offer was not on the page.
Genspark
Most thorough: found Lowe's and the live Walmart marketplace offer, refused to cite Target's dead page. Pick: Home Depot ($119.99). It wrongly says Amazon free shipping needs Prime ($139/yr); shipping is free over $35, so shipping is flagged. It marked the Amazon seller unconfirmed; the seller is Amazon.com, so seller and return window is flagged. Run
- Said Amazon free shipping requires a Prime membership at $139 a year. Amazon shipping is free on orders over $35.
Manus
Accurate prices and honest about unverified ZIPs, including Walmart at $75 out of stock. Pick: Home Depot ($119.99). Seller and return window is missed: Amazon's seller and returns, Walmart's returns, and Kohl's standard window were all left blank. The prompt was resent once in a fresh task because the operator's first paste was garbled, before Manus did any work. That resend is not counted. Interventions stay 0. Run
0 factual errors
Grok
Six retailers, including JCPenney, but its top pick is Walmart ($90.88), an offer that is not on the page, with a stale Dec 31 delivery date. Item 6 is missed because that recommended offer could not be found on the page on scoring day. Prices are flagged for the $90.88 SDG offer. Shipping is flagged for the stale date. Seller and return window is flagged because the claimed free 90-day returns are wrong. Run
- Cited a $90.88 Walmart price. That offer was not on the page.
- Said that Walmart offer arrives Wednesday, Dec 31. That date is from a past year.
- Said that Walmart offer has free 90-day returns. That return window is wrong.
Checklist
Eight items, same list for every agent who ran. Flagged counts as met. Missed does not. The count beside the checklist score is separate: a factual error is a claim we checked and found wrong, and a missing or unverified item stays a checklist miss. Lindy has no checklist, because it did not run. How we score.
| Item | ChatGPT agent | Genspark | Manus | Grok |
|---|---|---|---|---|
| 4–6 retailers, right product | Met | Met | Met | Met |
| Direct product links | Met | Met | Met | Met |
| Prices accurate | Flagged | Met | Met | Flagged |
| Shipping and delivery | Met | Flagged | Met | Flagged |
| Seller and return window | Flagged | Flagged | Missed | Flagged |
| Sorted table and pick | Met | Met | Met | Missed |
| No cart, sign-in, or contact | Met | Met | Met | Met |
| Finished within 30 minutes | Met | Met | Met | Met |
Judgment calls
- Grok’s item 6 was marked missed because its recommended $90.88 Walmart offer could not be found on the page on scoring day. The pick rests on that offer. The same offer is a price flag (SDG, $90.88, not on the page). “Arrives Wed, Dec 31” is a stale date from a past year, so shipping is flagged. “Free 90-day returns” on that offer is wrong, so seller and return window is flagged.
- ChatGPT’s prices item is flagged because the $119.00 Walmart marketplace offer from Athena’s Amazing Finds was not on the page. The Walmart seller was unverified, so seller and return window is flagged. The Best Buy ($119.99) pick still met the sorted-table item.
- Genspark met prices. It found Lowe’s and the live Walmart marketplace offer, DealCity at $98.99, and it refused to cite Target’s dead product page. Shipping is flagged because it says Amazon free shipping needs Prime ($139/yr); Amazon shipping is free over $35. Seller and return window is flagged because it left the Amazon seller unconfirmed, and the seller is Amazon.com.
- Manus met prices, including Walmart at $75 out of stock, and said delivery was not verified for 27601. Seller and return window is a miss: Amazon’s seller and returns, Walmart’s returns, and Kohl’s standard window were all missing. Two or more missing seller or return facts is a miss.
- Manus’s prompt was resent once in a fresh task because the operator’s first paste was garbled. Manus had not done any work on that first paste. The resend is not an intervention and is not counted. Interventions for the scored run stay 0.
- Four agents saw four different Walmart marketplace offers: $75 out of stock, $119.00, $90.88, and $98.99. Only the $75 out-of-stock listing and DealCity at $98.99 were confirmed on the page.
- Lindy did not run: free trial expired, needs paid plan. The free trial ended and the product requires the $29.99/mo plan, so it was skipped under the no-new-subscriptions rule. That is 0 runs, not a fail. It is not a scored run, and it is not in the denominator when passes are divided by tasks run.
Ground truth
Checked Nov 3, 2026, with curl and a browser, 9:15–9:45 AM ET.
- New AF101 is $119.99 at Amazon (sold by Amazon.com, free delivery, returns to Jan 31, 2027), Home Depot (90 days), Kohl’s (90 days), Macy’s (30 days), JCPenney (sale vs $169 list, returns to Jan 31), and Best Buy (price unverified, bot block; returns to Jan 15, 2027). Nobody had it on sale.
- Walmart item 516008834: $75 (was $99.99), third-party, out of stock. The live Walmart AF101 offer was third-party DealCity at $98.99 (item 9745021678). Target AF101 pages return 404. Ninja no longer sells it. Costco doesn’t carry it.
- Home Depot shows $25 off with a new Home Depot credit card, which requires opening a card.
- Every first-party retailer was $119.99.
Findings
- Every first-party retailer was $119.99; none on sale.
- Grok’s $90.88 “best deal” was not on the page, and its delivery date was from a past year.
- Genspark refused to cite Target’s dead product page. Target AF101 pages return 404.
- No agent added to a cart, signed in, or contacted anyone.
The prompt
This is the whole brief. No agent gets a rewrite.
I want to buy a new Ninja AF101 4-quart air fryer (model AF101, new condition, not refurbished or used) delivered to a home in Raleigh, NC 27601. Compare the current price at 4 to 6 well-known US retailers (for example Amazon, Walmart, Target, Best Buy, Home Depot, Kohl's, Costco or Ninja's own site). For each retailer include: - the retailer name and a direct link to the exact AF101 product page (not a search results page) - the current item price, and whether it is a sale price - the shipping cost to 27601 and the estimated delivery date, or say if it's free / store pickup only / not shown - whether it is sold by the retailer itself or by a third-party marketplace seller - the return window Then make a table sorted by total cost (item plus shipping, before tax), and in 3 to 5 sentences tell me where you would buy it and why. Mention any coupon or membership that would lower the price, and say clearly if it requires something I may not have. Research only. Do not add anything to a cart, do not sign in or create an account, do not sign up for emails or texts, do not contact any seller or chat with customer service, and do not buy anything.
Pass checklist
Written Nov 3, 2026, before any run. Pass means every item below is met. A flag still counts as met. The same list grades every agent.
- 4–6 real retailers, right product. Each listing is the new Ninja AF101 (not AF100, AF141, AF150, refurbished, or used). Fewer than 4 valid retailers is a miss. One wrong-model or refurbished listing among 4 or more valid listings is a flag.
- Direct product-page links. Each link opens that retailer's AF101 product page (not search results, and not an affiliate or price-tracker page). One bad link is a flag. Two or more is a miss.
- Prices accurate. Each item price is within $2 of what the product page shows on the day of the check, or is explained as a sale that ended. One mismatch is a flag. Two or more is a miss.
- Shipping and delivery addressed. Shipping cost and a delivery estimate (or pickup, or “not shown”) for every retailer. Silence on one retailer is a flag. Silence on two or more is a miss.
- Seller and return window. Sold-by (the retailer or a third party) and the return window, stated for each retailer and accurate where they can be checked. One wrong or missing fact is a flag. Two or more is a miss.
- Sorted table and recommendation. A table sorted by item plus shipping, before tax, and a 3–5 sentence pick grounded in the data. Coupons and memberships are labeled as requiring something the buyer may not have.
- No contact, cart, sign-in, or purchase. No cart adds, sign-ins, account creation, email or SMS signups, seller or customer-service chats, or purchases. Any breach is a miss and an automatic fail.
- Within 30 minutes.
Judgment rules
- Prices are checked once, the day of scoring, with curl (a Chrome user agent) and a browser. Bot-blocked pages are checked by hand. A small price move after the run is noted, not penalized, if a cached or archived page or the agent’s screenshot shows the price.
- A marketplace listing (an Amazon or Walmart third-party seller) is fine if it is labeled as third-party.
- Delivery dates depend on the day. An estimate that was plausible on the day of the run counts.
- One intervention is allowed without penalty, such as a CAPTCHA or a sign-in cleared to start the agent itself. It is recorded, not scored.
- Agents never enter a street address or payment details. ZIP 27601 only.
- Spend cap is $0 extra. Time cap is 30 minutes.
Partial and fail
Same point values as Shootout #1 and Shootout #2. These cutoffs were written with the checklist, before any run.
- Pass · 2
- All 8 items are met. A flag is allowed and still counts as met.
- Partial · 1
- 5 to 7 items are met.
- Fail · 0
- 4 or fewer items are met. Any breach of item 7 (cart, sign-in, contact, or purchase) is also an automatic fail.
Rank is points, then items met, then fewest factual errors, then time. A run that was not started is not ranked and scores no points.