Small talk is detected first
Keplar checks whether a message is plain social chat before it classifies anything else. The check is conservative and works on whole phrases: greetings, thanks, "how are you", goodbyes, "who are you", and short acknowledgements, up to 60 characters and 8 words.
Anything with a question, a request, digits, code or links, attachments or media, or a mode switched on is not small talk and goes through the full engine.
What small talk gets
- One free model, with a fallback to other free models and finally a static line.
- A short reply (capped at about 160 tokens).
- No panel, no verifier, no judge, no comparison section.
- No credits on any plan.
The reply is shown as plain text, without the consensus ring or the six sections.
Short factual questions
A short, plain lookup such as "What is the capital of Australia?" is classed simple and goes to one fast model. There is no review step, and the answer says "No cross-model verification needed for this question."
If the question needs an exact figure, a count, or arithmetic, it is not left to one model; see Exact checks.
When you want a second opinion on something simple
Choose Most thorough in the options next to the composer. Simple questions then get a panel too. Or rephrase the question as a decision ("Which of these is more likely and why?"), which the classifier reads as analysis.
Why this matters for credits
Casual messages cost nothing, and simple single-model answers are cheap, so the credits you pay for go to the questions where a panel helps. See Credits explained.
Related
- How Keplar understands a question: The classifier that decides task type, complexity and needed capabilities, what signals it uses, and why it is a transparent rule set rather than another model.
- Which questions use more models: A practical guide to what makes Keplar consult one model, three, or five: question type, length, stakes, forecasts, attachments, Deep Research and thoroughness, with examples.
- Credits explained: What a Keplar credit is, what costs more or less, how credits are counted from real provider charges, and what does not use credits.