I ran 95 returns through four models at once, not four times in a row
I run a small electronics and gadget shop in Johannesburg—phone cases, charging cables, Bluetooth speakers, power banks, the sort of thing people order on a Monday and return on a Thursday. We're on Takealot as a marketplace seller and run our own Shopify store for direct orders. About 70% of our revenue comes through Takealot, which means 70% of our return headaches do too. In the last week of February, during a stretch of stage-four load-shedding that knocked out our office three times a day, I had 95 return requests sitting in my inbox. Takealot gives buyers 30 days to return most items, and South African consumer law backs that up, so returns aren't optional. They're just Tuesday.
I'd been using Claude Sonnet 4.5 for return emails since January. It writes clearly, stays polite, and doesn't argue with customers. But I kept wondering whether Gemini, Llama, or DeepSeek might handle specific return reasons better. Damaged goods need empathy and speed. Change-of-mind returns need firmness without sounding cold. Wrong-item claims need precision because Takealot flags sellers who admit fault too often. Testing each model one at a time would take four weeks and four sets of customers. So I used Kryotta's Compare feature, which lets you send the same prompt to Claude, Gemini Pro, Llama 3.3 70B, and DeepSeek V3 at once and see all four responses side by side in about six seconds.
I didn't test every model on every return. I tested them in batches by return reason, tracked which responses I actually sent, and noted which ones led to escalations or follow-up complaints. This isn't lab science—I didn't control for customer mood or time of day—but it's real data from a real week of load-shedding chaos, and it changed how I route my prompts.
Damaged goods: Claude won 34 out of 38 times
Damaged-goods returns need two things: you have to sound genuinely sorry, and you have to tell the customer exactly what happens next. No fluff, no corporate language, no "we take quality seriously" paragraphs that say nothing. I tested all four models on 38 damaged-goods returns—cracked phone screens, crushed packaging, charging cables that arrived frayed. I gave them the same prompt every time: "Customer received [item] damaged. Write a return approval email that's warm, apologises clearly, explains the refund timeline, and offers a discount code for their next order. 80 words maximum."
Claude Sonnet 4.5 wrote the best responses 34 times. It apologised in the first sentence, confirmed the refund would process within 5–7 business days, and offered a 10% discount code without making it sound like a bribe. Gemini Pro came second—good tone, clear steps—but it added an extra sentence about "quality standards" in 12 responses that felt like filler. Llama 3.3 70B was too casual. It used phrases like "sorry about that" and "we'll sort it out", which worked for some customers but felt too informal when the item cost over R800. DeepSeek V3 wrote the longest responses and included a sentence asking the customer to send photos, which Takealot's return system already requires. That extra step annoyed three customers enough that they replied asking why I needed more proof.
I sent Claude's response 34 times, Gemini's four times. Token cost for Claude averaged R0.42 per email. Response time was 4–5 seconds even during load-shedding because I was working off my phone's hotspot and Compare cached the results.
Change of mind: Gemini Pro took 29 out of 31
Change-of-mind returns are the hardest to write. You have to approve the return because the law says so, but you also want the customer to feel slightly guilty so they think twice next time. Too stiff and they leave a bad review. Too friendly and they'll order three sizes of everything forever. I tested 31 change-of-mind returns with the same four models. Prompt: "Customer is returning [item] because they changed their mind. Approve the return politely, confirm the refund timeline, remind them of our return policy in one sentence, and stay professional but not cold. 70 words."
Gemini Pro won this category. It wrote responses that acknowledged the return, explained the 5–7 day refund window, and added a single line—"Our 30-day return policy covers change-of-mind returns for items in original packaging"—that reminded customers of the rule without sounding like a lecture. Claude was close, but it skipped the policy reminder in 18 responses, which meant I had to add it manually. Llama 3.3 70B was too warm. It thanked the customer for their "honesty" in one response, which made no sense. DeepSeek V3 added a sentence encouraging the customer to "browse our store again soon", which felt desperate.
I sent Gemini's response 29 times, Claude's twice. Gemini cost R0.38 per email and took 5 seconds in Compare. One customer replied to say the email was "clear and fair", which is the best outcome you can hope for in a change-of-mind return.
Wrong item shipped: Llama 3.3 70B handled 14 out of 18
Wrong-item claims are high stakes. If the customer's right, you have to fix it fast. If the customer's wrong—maybe they ordered the black case but swear they clicked blue—you have to correct them without calling them a liar, because Takealot will side with the buyer unless you have bulletproof evidence. I tested 18 wrong-item returns. Prompt: "Customer says we shipped the wrong item. Check the order details: they ordered [item A]. Approve the return, apologise, confirm we'll send the correct item, and explain the timeline. 90 words."
Llama 3.3 70B surprised me here. It wrote responses that apologised, confirmed the correct item would ship within two business days, and included the order number and SKU in a short table at the bottom of the email. That table made it easy for the customer to verify what they'd actually ordered, and in four cases the customer replied to say "sorry, I was wrong". Llama's tone was warmer than Claude's but more precise than Gemini's. Claude wrote good responses but left out the order number twice, which led to follow-up questions. Gemini added a sentence about "reviewing our packing process", which sounded defensive. DeepSeek wrote the longest responses and included two paragraphs about our quality checks, which no customer cares about when they just want the right item.
I sent Llama's response 14 times, Claude's four times. Llama cost R0.36 per email in Compare and took 5–6 seconds. Two customers replied to thank me for the clear explanation, and one left a four-star review that mentioned "helpful customer service".
Delivery delays: DeepSeek V3 won 8 out of 8
I only had eight delivery-delay returns in this batch, all from customers who'd waited more than 10 days and wanted to cancel. Takealot's delivery estimates are optimistic, and load-shedding slows down the couriers, so delays happen. Prompt: "Customer is returning [item] because delivery took too long. Approve the return, apologise for the delay, explain the refund timeline, and stay empathetic. 75 words."
DeepSeek V3 wrote the best responses here. It apologised in the first sentence, acknowledged that "delivery delays are frustrating, especially when you're counting on an item", and confirmed the refund without adding corporate language. The empathy felt real. Claude was close but used the phrase "we understand your frustration" in six responses, which sounds like a script. Gemini wrote clear emails but didn't acknowledge the delay directly—it just said "your refund will process within 5–7 days", which felt cold. Llama added a sentence offering a discount code, which was nice but felt like I was trying to buy forgiveness.
I sent DeepSeek's response all eight times. Cost was R0.44 per email, the highest of the four models, but the responses were worth it. One customer replied to say "I appreciate the honesty", and none of the eight escalated to Takealot support.
Compare saved me 11 hours and R340 in wasted API calls
Before I started using Compare, I would have tested these models one at a time over four weeks, sending real emails to real customers and hoping I didn't pick the wrong model for the wrong return reason. Compare let me see all four responses in one screen, pick the best one, and send it in under 20 seconds per return. Total time for 95 returns: about 30 minutes of active work spread across three days, plus another hour reviewing responses and tracking which models I'd chosen.
If I'd tested sequentially, I'd have spent 11 hours writing, comparing, and re-testing. If I'd stuck with one model for everything, I'd have sent 34 mediocre change-of-mind emails using Claude, or 38 too-casual damaged-goods emails using Llama. Compare cost me R1.60 per batch of four responses—R0.40 per model—which added up to R152 for the full 95 returns. Running each model separately would have cost R228 because I'd have generated responses I never used.
The other benefit: I learned which model to route to for each return reason. Now I use Claude for damaged goods, Gemini for change of mind, Llama for wrong-item claims, and DeepSeek for delays. I don't need Compare for every email anymore, but I use it when a return doesn't fit a clear category or when I'm too tired to guess.
Questions people ask
Can you use Compare for non-English return emails?
Yes. I tested a few Afrikaans returns using the same four models and Claude handled them best, but my sample size was only six emails. Gemini and Llama both mixed in English phrases twice.
Does Compare work during load-shedding?
It works fine on a phone hotspot. I ran most of these comparisons on 4G during stage-four cuts and the responses came back in 5–7 seconds. The Compare screen caches results, so if your connection drops after the models respond, you don't lose anything.
Which model is cheapest for high-volume returns?
Llama 3.3 70B and Gemini Pro both averaged R0.36–R0.38 per email in my tests. Claude was R0.42, DeepSeek R0.44. If you're processing 200 returns a month, the difference is about R16, so I'd pick based on quality, not cost.
Do you need a separate Kryotta account for each model?
No. Compare runs all four models under one workspace. You pay per token across whichever models you use, and you can see the cost breakdown per comparison in the usage panel.
I'm still testing Compare on product descriptions and WhatsApp Business replies, but for returns it's already changed how I work. If you're managing Takealot customer service on your own and you're tired of guessing which model to use, try Compare on your next ten returns and see which responses you actually send. You'll know in an hour whether it's worth the R4 per comparison. Start a free Compare session here and bring your own return emails—you'll see all four responses before you pay anything.
