Myth one: newer models always win Razorpay disputes faster
I manage customer service for a Jaipur-based jewellery store—silver earrings, kundan sets, wedding pieces—that does ₹12 lakh a month through Shopify and takes most payments via Razorpay. In January we had 47 disputes in three weeks: "item not received", "wrong design shipped", "UPI payment deducted twice", and the classic "my husband ordered this without asking me". Razorpay gives you five days to respond with tracking proof, screenshots, and a coherent explanation. Win and the ₹3,800 stays in your account. Lose and you've shipped a necklace you'll never see again.
I'd been using GPT-4o since November because everyone said it was the best. When Gemini 2.0 launched in December, I assumed I should switch. Both are new, both claim better reasoning, both sound impressive in the release notes. I tested them on ten disputes each, then compared the results to what I got from older models—Claude Sonnet 4.5 and Gemini Flash—that I'd ignored because they felt outdated.
GPT-4o and Gemini 2.0 both wrote detailed, structured responses. They pulled in tracking numbers correctly, referenced our return policy, and formatted everything clearly. But they took 8–12 seconds per response and cost me ₹2.40 per dispute on average when I tracked token usage. Claude Sonnet 4.5 gave me the same structure in 4 seconds and cost ₹1.80. Gemini Flash—the model I'd written off as too basic—handled 80% of disputes perfectly in 2 seconds for ₹0.60 each. The Razorpay team doesn't care whether your response came from the newest model. They care whether you included the AWB number, explained the UPI deduction, and attached the delivery screenshot. Flash did all three.
I still use GPT-4o for the 20% of disputes that need more reasoning—cases where a customer claims they received a broken item but our courier photos show it intact, or when someone disputes a custom order they approved over WhatsApp. But for straightforward "where's my order" disputes, Flash wins on speed and cost every time.
Myth two: flash models can't handle UPI double-deduction errors
UPI double-deduction errors are the worst kind of dispute. The customer sees two debits on their bank statement, panics, files a Razorpay dispute, and expects you to fix it immediately even though the second deduction usually reverses within 48 hours. You need a response that explains what happened, confirms you only received one payment, tells them when the refund will hit their account, and doesn't sound like you're blaming them for their bank's delay.
I thought this needed a reasoning-heavy model—GPT-4o or Gemini 2.0—because the explanation has to cover UPI settlement flows, Razorpay's reconciliation process, and your store's refund policy without confusing someone who's already annoyed. I tested both on five UPI disputes in early February. Both gave clear, accurate explanations. Both took 10+ seconds. Both cost ₹2–2.50 per response.
Then I tried Gemini Flash with a tighter prompt: "Customer sees two UPI debits for order #4721. Only one payment reached us. Explain why the second debit will reverse, when they'll see the refund, and what we've done to confirm they weren't charged twice. Keep it under 120 words." Flash returned a perfect response in 2 seconds for ₹0.50. It covered the settlement delay, confirmed we'd checked our Razorpay dashboard, and gave a timeline. The customer replied "okay, I'll wait" and closed the dispute.
Flash models aren't too simple for UPI errors. They just need a specific prompt. If you ask for "a response to this UPI dispute" you'll get generic waffle. If you tell the model exactly what to explain and how long the answer should be, Flash handles it faster and cheaper than anything else.
Myth three: you need the latest model for GST invoice generation
GST invoices are boring, repetitive, and rule-bound. You need the customer's GSTIN, your own GSTIN, the HSN code for each product, the taxable amount, CGST, SGST or IGST depending on the state, and the total. The format is fixed. The rules don't change. This is exactly the kind of task that doesn't need a frontier model.
I tested GPT-4o, Gemini 2.0, Claude Sonnet 4.5, and Mistral Large on 20 invoices in January. All four got the format right. All four calculated the tax correctly. GPT-4o and Gemini 2.0 took 6–8 seconds per invoice and cost ₹1.80–2.20. Claude Sonnet 4.5 took 4 seconds and cost ₹1.50. Mistral Large took 3 seconds and cost ₹0.90.
Then I tried Amazon Nova Lite, which Kryotta added in December. It's the smallest model in the Nova family, designed for simple structured tasks. I gave it a prompt with the customer's details, the product list, and the tax rates, and asked for a GST-compliant invoice in plain text. It returned a perfect invoice in under 2 seconds for ₹0.40. I've generated 140 invoices with Nova Lite since mid-January and haven't had a single error.
If you're using GPT-4o or Gemini 2.0 for invoice generation, you're paying five times more than you need to for a task that doesn't require reasoning. Use Nova Lite or Mistral Large and save the expensive models for the work that actually needs them.
Myth four: you can't automate refund emails during Diwali without losing your tone
Diwali is chaos. Order volume triples, delivery delays spike, and refund requests flood your inbox—some legitimate, some speculative, all urgent. I used to write every refund email manually because I was convinced automation would make us sound like a bot. In October I had 83 refund requests in nine days. I couldn't keep up.
I tested three models—Claude Sonnet 4.5, Gemini Flash, and Llama 3.3 70B—on ten refund emails each, using the same prompt: "Write a refund email for [reason]. Acknowledge the issue, confirm the refund amount and timeline, apologize without groveling, and offer a 10% discount on their next order. Keep it warm and human." I wanted to see which model could handle the tone without sounding corporate or overly apologetic.
Claude Sonnet 4.5 nailed it. The emails felt like something I'd write—direct, warm, no fluff. "We're sorry the mehendi cones arrived dried out. We've processed a full refund of ₹890, and you'll see it in your account within 5–7 working days. Here's a 10% discount code for your next order: DIWALI10. Let us know if you need anything else." It took 4 seconds and cost ₹1.20 per email.
Gemini Flash was faster—2 seconds, ₹0.50—but the tone was slightly off. It used phrases like "we sincerely regret the inconvenience" and "please accept our apologies", which felt stiff. Llama 3.3 70B was good but verbose. It added unnecessary sentences about our "commitment to quality" that I had to delete.
I used Claude Sonnet 4.5 for all 83 refunds. I edited maybe 15% of them—adding a line about a specific product issue or adjusting the discount offer—but the rest went out as drafted. Three customers replied to say they appreciated the quick response. No one accused us of sounding like a bot.
Myth five: Compare arena is overkill for a small store's dispute budget
When I first saw Kryotta's Compare feature—type one prompt, get responses from four models side-by-side—I assumed it was for people running tests or writing long-form content. I didn't think it made sense for dispute emails, where you just need one good answer fast.
I tried it in February on a complicated Razorpay dispute. A customer claimed they'd ordered a gold-plated necklace but received a silver one. Our Shopify order log showed they'd selected silver. Our WhatsApp chat showed they'd confirmed silver. But their dispute email insisted they'd wanted gold and we'd made a mistake. I needed a response that presented the evidence clearly without sounding accusatory.
I ran the prompt through Compare with Claude Sonnet 4.5, Gemini Flash, GPT-OSS 120B, and Mistral Large. Claude's response was firm and evidence-focused but felt slightly cold. Gemini's was warmer but buried the key evidence in the second paragraph. GPT-OSS struck the right balance—it opened with empathy, laid out the order log and WhatsApp screenshots in sequence, and closed with an offer to exchange the item if they covered return shipping. I used GPT-OSS's draft with one sentence edited. We won the dispute.
Compare didn't add time—it took 6 seconds to generate four responses, and I spent maybe 30 seconds reading them. But it gave me options. I could see which model handled the evidence best, which one sounded too defensive, and which one got the tone right. For disputes where ₹3,000–8,000 is on the line, seeing four approaches before you hit send is worth it.
I now use Compare for any dispute that feels borderline—cases where the customer has partial evidence, or where our records conflict with their claim, or where the tone matters as much as the facts. For straightforward disputes I still use Gemini Flash on its own. But Compare has saved me from losing at least three disputes I would've handled badly if I'd only seen one model's response.
Questions people ask
Which model should I use for basic Razorpay disputes where the tracking shows delivery?
Gemini Flash. It's fast, cheap (₹0.50–0.70 per response), and handles straightforward evidence-based emails perfectly. Save the expensive models for disputes that need more reasoning.
Does GPT-4o ever make sense for e-commerce support?
Yes, for the 15–20% of disputes where the facts are messy—conflicting delivery photos, custom orders with verbal approval, or cases where you need to explain a complex refund policy. But not for routine "where's my order" emails.
Can I use Nova Lite for customer service emails or just invoices?
Just invoices and other structured tasks—order confirmations, shipping updates, GST summaries. For emails where tone matters, use Claude Sonnet 4.5 or Gemini Flash.
Is Compare worth using if I only handle 10–15 disputes a month?
Yes, especially for high-value disputes. Losing one ₹5,000 case because your email sounded defensive costs more than a month of Kryotta. Compare lets you see four approaches in six seconds.
I've been using Kryotta since December to handle Razorpay disputes, GST invoices, and refund emails during peak season. The Compare feature and access to six models in one workspace—Claude, Gemini, Llama, Mistral, DeepSeek, and the Nova family—means I can match the right model to each task without paying for five subscriptions. If you're still using one model for everything, try Kryotta and see what you're overspending on.
