Skip to main content

Choosing a model

Two models are available for your chatbot.

ModelCharacter
gpt-4o-mini (default)Fast and inexpensive. Handles the great majority of customer questions well.
gpt-4oSlower and more expensive. Better at multi-part questions, subtle comparisons and careful reasoning over long passages.

Start with the default​

gpt-4o-mini is the default for a reason. For the work most chatbots do — answering a question from a retrieved passage of your own content — the more expensive model produces a very similar answer.

Remember that the hard part is retrieval, not generation. If the right passage is found, both models answer it well. If it is not found, neither model can help, because both are forbidden from inventing one. Upgrading the model does not fix a knowledge base problem.

When gpt-4o earns its cost​

Switch if your customers routinely ask questions like these:

Multi-part questions.

"Do you deliver to Aberdeen, and if so how long does it take and is there a minimum order?"

Comparisons across passages.

"What's the difference between the Standard and Professional plans for a team of five?"

Reasoning over conditions.

"I bought it 40 days ago but only opened it last week — can I still return it?"

Long, dense source material. Technical documentation, legal or regulatory text, detailed specifications.

Trying it properly​

Do not switch on a hunch. Test it.

  1. Collect five to ten real questions from your live inbox — ideally ones where the answer was weak
  2. Ask them on gpt-4o-mini in the dashboard Chat page and save the answers
  3. Switch to gpt-4o and ask exactly the same questions
  4. Compare

If the answers are noticeably better, keep it. If they are barely different — and often they will be — switch back and put the effort into your content instead.

Cost and speed​

gpt-4o costs meaningfully more per message and takes longer to respond. Both matter at volume: a slower first word is a real difference in a chat widget, where visitors are watching the reply appear.

Because answers stream word by word, the perceived gap is smaller than the raw timing suggests. But it is there.

Changing the setting​

Chatbot Settings → model. The change applies to the next message; existing conversations are unaffected and nothing needs retraining.

You can set this per chatbot, so a straightforward FAQ bot and a technical support bot on the same account can use different models.