Which AI model for a customer service chatbot?
A chatbot that answers customers: order tracking, FAQs, returns, with a human handover.
GLM-5.3 Flash
- Index
- 153.2
- Price / M tokens
- €0.13 / €0.44
- Cost / 1 M
- ≈ €466
The cheapest model on the Pareto line: each message rereads the history, so the price per message matters.
See in the ranking →Qwen3.8 27B
- Index
- 151.6
- Price / M tokens
- Free (local)
Open weights under the Apache 2.0 license: conversations and customer data stay on your server.
See in the ranking →Claude Sonnet 5.5
- Index
- 164.4
- Price / M tokens
- €1.78 / €8.89
- Cost / 1 M
- ≈ €7,109
On the Pareto line, with one of the best capability scores: more accurate answers and better tone with unhappy customers.
See in the ranking →Why these choices
For a chatbot, the model matters less than its guardrails: staying in scope, resisting injection attempts, promising nothing and handing over to a human. Test it with attacks before going live, and disclose that it is an AI (EU AI Act, Article 50).
Take action
Red teaming: 12 attacks on a customer service chatbot
12 attacks to replay on a U.S. e-commerce support chatbot: impersonation, data leaks, prompt injection, unauthorized refunds. Expected reply for each.
Minimal private AI chat: Ollama and Open WebUI
The simplest stack for a private AI chat: Ollama and Open WebUI with docker-compose, interface on port 3000.
Guardrail: mask U.S. personal data (SSN, phone, card)
Guardrail that masks U.S. personal data in AI assistant replies: SSNs, +1 phone numbers, personal emails and card numbers. Rules, 12 tests, code.
Data: Epoch AI index (CC BY 4.0), developers' official prices converted at the ECB rate, updated 2026-10-19 · AI model ranking: intelligence, benchmarks and price