Response controls
Three settings shape how a model answers: the profile controls writing style, effort controls length, and thinking controls whether the model reasons before replying.
Profile and effort live behind the three-dot button at the left of the composer. Thinking settings are in Settings → Thinking.
Profiles
A profile changes how adventurous the model is when picking each next word. Lower settings produce steadier, more predictable text; higher settings produce more variety, and more mistakes.
| Profile | Character | Use it for |
|---|---|---|
| Casual | Balanced. Varied phrasing, conversational rhythm. | General chat. This is the default and the right choice most of the time. |
| Logical | Tightly constrained. Picks the likeliest wording, keeps answers short. | Factual questions, definitions, arithmetic, anything where you want the same answer twice. |
| Creative | Unconstrained. Surprising word choices and much more variation. | Brainstorming, fiction, wordplay. Expect more errors — this setting trades accuracy for range. |
Creative widens sampling substantially. That is exactly what you want for fiction and exactly what you do not want for facts — a small model on a wide setting will invent details fluently. Switch to Logical before asking anything you intend to rely on.
Effort
Effort controls how long a reply is allowed to run. Standard uses the profile's normal budget. Extended raises it by roughly 85 percent, giving the model room to work through something properly instead of stopping mid-thought.
Extended is worth turning on for explanations, comparisons, and step-by-step work. It is not worth it for short factual questions, where the only effect is a slower reply.
Haiku H2 has a 1,000-token context window, which does not leave room for an extended budget. On those models the effort setting resolves to Standard automatically. Your choice is remembered and comes back when you return to an Epos model — it is not being ignored, just not applicable.
Thinking
Some replies are preceded by a reasoning pass, shown as an expandable Thought for… line. Normally the model decides when this is warranted. Two settings override that, both in Settings.
Force thinking
Makes the model reason before every reply rather than only when it judges reasoning useful. This helps on multi-step problems. On simple conversational turns it produces a stilted deliberation about a question that did not need one — which is why it is off by default.
Augmented thinking
Marked experimental, and requires force thinking to be on. It analyses your prompt for intent and topic first, and seeds the model's reasoning with that analysis instead of letting it start cold.
Experimental means what it says: it can sharpen a reply and it can send the model down a confidently wrong path if the analysis misreads what you were asking. Try it, keep an eye on the results, and turn it off if replies get worse.
Putting them together
| You want | Settings |
|---|---|
| A quick factual answer | Logical, Standard, thinking off |
| Something explained properly | Casual, Extended |
| A multi-step problem worked through | Logical, Extended, force thinking on |
| Ideas and variety | Creative, Extended |
| Ordinary conversation | Casual, Standard — the defaults |
These are starting points rather than rules. If replies are too short, raise effort. If they wander or invent things, move toward Logical. If they feel mechanical, move toward Casual.