We ran 60 real support tickets through GPT-5 and Claude Sonnet 4.6 with identical system prompts. Here's where each one wins, the exact prompts that perform, and the latency/cost tradeoffs that decide which one ships.
I get asked weekly which model to use for customer support. The "which is smarter" question is the wrong one. The right one is: which one ships the right answer in under 4 seconds, costs under $0.04 per resolved ticket, and doesn't hallucinate a refund policy that doesn't exist? 60 real anonymized support tickets from three TrainYourAgent customers (a DTC skincare brand, a SaaS onboarding team, a roofing scheduling desk) through GPT-5 and Claude Sonnet 4.6 with identical system prompts, identical RAG context, identical evaluator rubric. The test setup Three test buckets, 20 tickets each: Bucket A — Policy questions: refunds, shipping windows, return eligibility, plan upgrades/downgrades.
60 real anonymized support tickets from three TrainYourAgent customers (a DTC skincare brand, a SaaS onboarding team, a roofing scheduling desk) through GPT-5 and Claude Sonnet 4.6 with identical system prompts, identical RAG context, identical evaluator rubric. The test setup Three test buckets, 20 tickets each: Bucket A — Policy questions: refunds, shipping windows, return eligibility, plan upgrades/downgrades. Bucket B — Multi-turn troubleshooting: install error, billing dispute, account recovery, "this feature isn't working." Bucket C — Empathy + escalation: angry customer, urgent, billing emergency, "I want to speak to a manager." Same system prompt for both: Same RAG: top-5 chunks from each brand's actual KB at retrieval temperature 0.0. Same temperature on the models: 0.3. Evaluator: a third LLM (Claude Opus 4) blind-scoring on (a) factual accuracy vs the KB, (b) appropriate empathy, (c) brevity, (d) safe escalation behavior.
We ran 60 real support tickets through GPT-5 and Claude Sonnet 4.6 with identical system prompts. Here's where each one wins, the exact prompts that perform, and the latency/cost tradeoffs that decide which one ships. It is filed under AI Strategy because that is where operators looking for this problem actually start, and it is written from production work rather than from a content calendar.