Which AI Model Powers Your Website Chatbot, and Does It Matter?

The AI model behind a website chatbot matters less than most people think. For a support bot, answer quality depends mostly on your content and on how well the tool finds the right passage to answer from. The model decides the rest: how reliably it follows rules, whether it admits when it doesn't know, how fast it replies, how well it handles other languages, and what each reply costs.

I build ChatBento, which runs on Anthropic's Claude Haiku 5.5, one of the smallest and fastest models available. Below is why I chose it, where a bigger model would do better, and what to ask any chatbot vendor about the model behind their answers.

How does a website chatbot answer a question?

In two steps:

1. Find. The tool searches your pages and Q&As for the passages that best match the visitor's question. This is retrieval, and it's done by the chatbot tool, not by the model whose name is on the box.
2. Write. Those passages, the bot's rules and the question go to a large language model, which writes the reply.

If step one hands the model the wrong passage, or nothing because the answer isn't on your site, no model can write the right answer. The best it can do is say it doesn't know.

Anthropic's own research shows how much the finding step matters. In a 2024 write-up on contextual retrieval, changing how documents were indexed and searched cut failed retrievals by 49%, and by 67% with an extra re-ranking step. Anthropic says better retrieval carries through to better answers. None of that involved changing the model.

What a better model won't fix

Most wrong answers I see on small business sites come from the content, not the model:

• Missing content. The answer isn't on the site, so the bot has nothing to work from.
• Stale pages. An old post still lists last year's prices, and the bot quotes it.
• Conflicting pages. The shipping page says 3 days and the FAQ says 5.
• Vague questions. "How much is it?" on a site with twelve products.

Swapping to a more expensive model fixes none of these. Adding a Q&A or deleting an outdated page does. I went through each cause and its fix in why AI chatbots make things up.

What does the model actually change?

Once the right passage is in front of it, the model still decides five things that matter for support.

Following the rules

A support bot runs on rules: answer only from this content, don't guess prices, decline unrelated questions, hand off when someone asks for a person. Models differ in how consistently they keep to those rules over a long conversation, especially when a visitor pushes back. For support this is the most important model trait, and you can test it in ten minutes.

Saying "I don't know"

Closely related. A model that prefers a fluent guess to an honest gap is a liability on a page about your refund policy. You want one that says it doesn't have the information and offers a way to reach you.

Speed

Visitors are watching a typing indicator. Smaller models reply faster, and shorter replies arrive sooner. Anthropic's models overview lists Haiku 5.5 as the fastest model in its current lineup.

Languages

Small models tend to lose more in less common languages. Anthropic publishes multilingual scores relative to English for an earlier pair of its models. The smaller Claude Haiku 4.5 scored 96.4% in Spanish, close to the larger Claude Sonnet 4.5's 98.2%. In Yoruba, Haiku 4.5 scored 52.7% against Sonnet 4.5's 79.7%. If your visitors write in widely spoken languages, a small model is close to a large one. If they don't, test in their language before you trust any model. I covered this in more depth in a multilingual website chatbot.

Cost per reply

This is where models differ most, and it ends up in your price.

Small fast models vs large ones: which suits support?

Here are Anthropic's published API prices for its four current models, checked November 2, 2026, with my own rough estimate of what one support reply costs. The estimate assumes about 4,000 tokens in (the bot's rules, the retrieved passages and the conversation so far) and 200 tokens out. Real numbers vary with the tool and the question.

Haiku 5.5's price is for prompts up to 100,000 tokens, which covers a normal support reply. All four are listed with the same 1 million token context window.

The most expensive model costs about 100 times more per reply. Anthropic describes Fable 5.1 as built "for demanding reasoning and long-horizon agentic work," and Haiku 5.5 as built for high-volume, latency-sensitive tasks. A support chat is the second kind: most questions are short lookups like "do you deliver on Saturdays?", where the hard part is finding the right passage, not reasoning about it.

A bigger model earns its price when the bot has to reason across many long documents, troubleshoot complicated technical problems step by step, or take actions in other systems. If that's your site, pay for it. For opening hours, prices, policies and "do you offer X", it rarely changes the answer.

Which models do other chatbot tools use?

Many tools let you choose. Here's what four vendors say on their own sites, checked November 2, 2026:

• SiteGPT offers GPT-4.1-mini and GPT-4.1, and says GPT-4.1 messages cost 10 times as much of your quota. Its Starter plan is up to 4,000 GPT-4.1-mini messages or 400 GPT-4.1 messages a month, or any mix.
• Chatling lists 32 models from several providers, costing from half a credit to 25 credits per AI message. Claude Haiku 5.5 is 1 credit; Claude Opus 5.5 is 10.
• Chatbase lists advanced models from OpenAI, Anthropic, Gemini, DeepSeek, MoonshotAI, Mistral and Z.ai as included on its Hobby, Basic and Business plans.
• Intercom's Fin runs on Intercom's own proprietary models, called Apex, which Intercom says are custom trained on customer service interactions.

A model menu is useful if you know what you're testing for. It also moves the decision, and the cost, to you: pick a bigger model and the same plan buys fewer replies.

Why ChatBento uses Claude Haiku 5.5

Every ChatBento reply is written by Claude Haiku 5.5 (model id claude-haiku-5-5), which Anthropic released in October 2026. My reasons:

• It's fast. Visitors get an answer while they're still on the page, which is the whole point of a chat widget.
• It follows instructions. ChatBento's built-in rules tell the bot to answer only from your content, never guess prices, policies or dates, decline unrelated questions, and reply in the visitor's language. In my testing, Haiku 5.5 keeps to them well, including when a visitor pushes.
• It's cheap enough for a flat price. Look at the table again: a small model is what makes 2,000 replies for $8 a month possible, and a free plan with 50 messages. On the most expensive model, my estimate for the AI alone is about $100, more than ten times the plan's price.
• Short answers suit it. ChatBento's built-in style asks for one to four sentences, or a short list when listing steps or options, with no preamble. Short replies are faster, cheaper and easier to read on a phone.
• Your data isn't used for training. Anthropic's privacy center says that by default it doesn't use inputs or outputs from its commercial products to train its models.

The trade-off is real: there's no bigger model to switch to inside ChatBento. If your bot needs deep reasoning over long technical documents, a tool that lets you choose a larger model may serve you better.

What should you ask a chatbot vendor about its model?

1. Which model writes the answers? A vendor should be able to tell you. "Our proprietary AI" is a fine answer if they also explain what it's good at.
2. Can it change without telling me? Models get retired, so every vendor switches eventually. Anthropic's deprecation policy promises at least 60 days' notice before retiring a publicly released model, and requests to a retired model fail. Claude Haiku 3.5, for example, was retired on February 19, 2026. Ask whether the vendor tells customers when they switch, and whether you can test first.
3. Is my data used for training? Ask about both the model provider and the vendor itself.
4. Does a bigger model use more of my allowance? On credit-based plans it usually does, so the price per month only means something once you know the model.
5. What does it do when it doesn't know? Ask, then check. My 30 questions to ask before launch include several your site can't answer, which is the fastest way to see how a model behaves at the edge of its knowledge.

The short version

Pick a chatbot for how well it answers from your content, and treat the model as one ingredient. Ask which model it uses, whether that can change, and whether your conversations train anyone's models. Then test it on your own questions, because that tells you more than any model name.

If you'd like to try a bot built on Claude Haiku 5.5, ChatBento has a free plan with 50 AI messages a month, and the paid plans are on the pricing page. For the basics behind all this, see the glossary entry on retrieval-augmented generation.