Models & reasoning effort

Pick a model per chat, tune reasoning effort, and run multiple models in parallel on the same folder.

Every chat in Dices runs a specific model at a chosen reasoning effort. You set these per chat, fall back to sensible defaults, and can even run several models side by side on the same project at once.

The model picker

Each chat has a model picker. The list of available models depends on which agent drives the chat:

Claude — Opus 4.8, Sonnet 4.6, Haiku 4.5, Fable 5.
Gemini — 2.5 Pro, 2.5 Flash, and others.
Other ACP agents surface whatever models that provider exposes.

Switch the model whenever you like: lean on a smaller, faster model for routine edits, and a larger model for hard reasoning or tricky refactors.

Tip: Opus 4.8 for the gnarly stuff, Sonnet 4.6 or Haiku 4.5 for fast iteration. You don’t have to commit — change the model mid-project as the work changes.

Reasoning effort

The reasoning effort control tells the model how hard to think before acting. The levels are:

Level	When to use it
low	Quick, mechanical changes where speed matters most.
medium	Everyday coding — a balanced default.
high	Multi-step changes, debugging, careful refactors.
xhigh	Hard problems that need extended deliberation.
max	The most demanding reasoning, when correctness beats speed.

Effort is provider-aware — it only does something where the model supports it:

Claude models support reasoning effort.
Haiku 4.5 and Sonnet 4.5 ignore it.
ACP agents don’t expose it at all — see Agents. For those, the effort control has no effect.

Note: Higher effort means more thinking time and token use. Reach for high+ when a task is genuinely hard; keep it at medium for ordinary work.

Multi-model per folder

Dices can run more than one model on the same folder at the same time, in parallel — and you set it up in plain language. Just ask the agent:

“Use Sonnet for this.”
“Let Gemini write the tests.”

When you do, Dices uses its open_chat tool to automatically open a NEW chat in the same folder with the requested model, running in parallel with your current chat. Each chat is isolated (see Git & worktrees), so the parallel models don’t collide on each other’s changes.

This makes it natural to, say, have Opus drive the main feature while Gemini writes tests alongside it — both in the same project, both at once.

Example

You:  Let Gemini write the integration tests for this module while you
      keep building the API.

Dices: (opens a new parallel chat in this folder running Gemini, and
        continues the current chat on the API)

Default model

You don’t have to choose every time. Set a default model:

Globally — applies to every new chat by default.
Per project — override the global default for a specific folder.

New chats start on the applicable default; the per-chat picker still lets you change it any time.

Agents — which agents and providers are available.
Permissions — how much each chat can do on its own.
Git & worktrees — how parallel chats stay isolated.