Every chatbot vendor's pricing page now names the models it runs on, and every buyer asks the same question: which one is best? For a chatbot that answers questions from your own website content, the honest answer is that the model is the third most important decision you will make, behind what you train it on and how it hands off. But it is still a decision, so here is how to think about it.
The landscape in 2026
The models behind website chatbots come from a handful of labs:
- OpenAI (the GPT family, which powers ChatGPT)
- Anthropic (the Claude family)
- Google (the Gemini family)
- Alibaba (Qwen), Zhipu AI (GLM), Moonshot AI (Kimi), and MiniMax, several of which publish open-weight models that anyone can host
Each lab ships models at several sizes, from small and fast to large and deliberate. The differences between labs at the same size are real but smaller than the differences between sizes within one lab. A small fast model from any of them will answer a shipping question correctly if the shipping page is in front of it; a large model from any of them will do a better job of synthesising an answer across three long documents.
What actually differs
Speed. Smaller models answer in a second or two; larger ones take longer. On a support widget, speed is a feature. Visitors wait for a fast answer and abandon a slow one.
Reasoning depth. Larger and reasoning-oriented models are better at multi-step questions: "if I return a sale item bought with a gift card, what do I get back?" For single-fact questions the difference is invisible. The reasoning models guide covers when depth pays.
Instruction following. Some models are better at staying inside the rules you give them, such as "answer only from the sources" and "hand off when unsure". This matters more than raw intelligence for a customer-facing bot.
Languages. All the major families handle the widely spoken languages well. If your customers write in less common languages, test the specific model on real questions before deciding.
Cost. Larger models cost more per message to run, which is why hosted platforms offer tiers. Open-weight models from Qwen, GLM, Kimi, and MiniMax have pushed the cost of capable inference down, which is part of why flat monthly plans with generous message allowances are now normal. The open-weight models article explains that shift.
What matters more than the model
Grounding. A chatbot that answers from your pages and says "not covered" when it should is trustworthy on any competent model. A chatbot answering from general knowledge is untrustworthy on the best model in the world, because it will describe a policy you do not have. This is the single biggest quality lever, and it is about setup, not the model. See training on your own data.
Source quality. Specific pages produce specific answers. The model cannot improve a vague FAQ.
The handoff. Every chatbot meets a question it cannot answer. What happens next is a configuration choice and it shapes how customers feel about the whole thing.
Testing in the customer's words. Teams test with their own vocabulary and then blame the model when customers get worse answers.
How to choose a tier
Hosted platforms package models into tiers rather than making you pick a vendor. Cortexvia's run from Mini to Super; the models page lists what each is for. A practical approach:
- Start on the default tier with your real sources.
- Ask your twenty most common questions, in customer phrasing.
- If single-fact answers are right and fast, stay there. That describes most support and FAQ chatbots.
- If multi-document or multi-step questions come back thin, step up a tier and re-test the same twenty.
- Pay for the tier that answers your questions, not the one with the most impressive name.
Frequently asked questions
Does Cortexvia use ChatGPT or Claude? Cortexvia offers model tiers rather than exposing a single vendor, and the tiers are described by what they are good at on the models page. For a grounded website chatbot, the tier decision is about speed and depth, and the grounding does the rest.
Should I pick the biggest model to be safe? No. It is slower and costs more, and for single-fact questions from your own pages it answers no better than a mid-sized model. Pick the smallest tier that passes your twenty-question test.
Will the model change under me? Labs retire and replace models regularly; hosted platforms handle that migration. What stays constant is your content, which is another reason to invest there first.