AI & answers
New AI model: Claude Haiku 5.5 for every bot
From today, every Mignuti bot answers with Claude Haiku 5.5. We put both models through the same questions: faster, and more careful with numbers.
Contents
From today, every Mignuti bot answers your visitors with Claude Haiku 5.5, Anthropic’s newest fast model. Until now we ran Claude Haiku 4.5. You don’t need to do anything: no setting, no new embed code, no restart. Your knowledge base, your tone and your rules stay exactly as they are.
A new model is not an automatic win. Before switching, we put both models head to head. Here is what came out of it, including the parts that did not get better.
How we tested
We took a bot we run ourselves: the website of a wedding photographer, with a pricing page, service descriptions, a blog and pages in German, English and Croatian. We asked 17 questions, first with the old model, then with the new one. And we used exactly the path a visitor takes: open the widget, ask, wait for the answer.
The set covered simple price questions, details about services and delivery times, questions the website has no answer to, questions in English, French and Croatian, a follow-up in the same conversation, an off-topic question and an attempt to make the bot reveal its instructions.
What got better
Faster answers. In the test, a complete answer was ready almost a quarter faster on average. For visitors that means less waiting until the full answer is there.
More careful with numbers. Asked “What does wedding photography for a full day cost?”, the old model quoted a price range as “typical in the region”. The pricing page, however, says “on request”. The new model stuck to what the page says, listed what is included and pointed to the contact form for a quote. That is how it should be: a bot that makes up prices is a liability for you.
Uses what is there. Asked “Do you also photograph weddings in Italy?”, the old model said it had no information. The website does say the photographer works all over Europe. The new model connected the two correctly: yes in principle, ask directly for details.
Honest about contradictions. For the delivery time of a wedding film, the website says four to eight weeks in one place and seven to 30 days in another. The old model picked one and left out the other. The new model gives both and says where each comes from. That doubles as a hint for you: this is a spot on your own website worth fixing.
What stayed the same
Answers in English and French were good and complete with both models. Phone numbers, scope of services and film length were right with both. Questions the website has no answer to were answered honestly by both, with “I don’t have information on that” and the way to get in touch, instead of something invented.
Our safeguards don’t depend on the model and still apply: the bot only answers from your sources, links are checked against those sources before they are shown, and private data in visitor messages is masked first. The attempt to extract the bot’s instructions ended the same way with both models, without ever reaching the model. The new model also brings its own safety checks. If it declines a request, our backup model takes over automatically, so your visitor is never left without an answer.
What did not get better
Both models shared two weaknesses, which means they are on us, not on the model:
- For “Do you have a drone with you?”, our own fact check held back a correct answer. The replacement sentence then came in English instead of German.
- For the Croatian question about the price of a couple shoot, the price was missing, although the Croatian pricing page is in the bot’s knowledge.
We are looking into both. We still write them down here, because switching models does not fix everything. A bot is as good as its sources and the checks around it.
What this means for you
Nothing you need to do. One thing is good to know: because the model changed, your bot generates answers to frequent questions afresh once instead of serving stored answers from the old model. If an answer sounds a little different from usual over the next few days, that is why.
If you spot an answer that could be better: you can review every conversation in the inbox, and under “Pending Questions” the “Add to knowledge” button gives the bot the right answer. Or write to us directly, we are always glad to hear about these.
Your own chatbot, built from your content
Add your website, upload PDFs or connect a folder. The bot answers only from your sources. Try it free for 7 days.