This page lists the AI models Octo uses for chat, images, web search, safety decisions and voice. The tables are generated from Octo’s OpenClaw config each time the docs are built.
Language Models
Role
Model
Notes
Primary
anthropic/claude-sonnet-4-6
Default for every agent without its own model
Fallback
anthropic/claude-haiku-4-5
Tried in order when the primary model is unavailable
Fallback
openai/gpt-6-astra
Tried in order when the primary model is unavailable
Image (vision)
anthropic/claude-sonnet-4-6
Used for image analysis
Image fallback
anthropic/claude-haiku-4-5
Web search
gemini
Provider behind the web search tool
Per-Agent Overrides
Agents listed here use their own model instead of the default. Decision models are covered in Decision Models below.
Agent
Model
Decision model
claude
anthropic/claude-opus-5-5
—
hassonic
openai/gpt-5.6-luna
typesafe/jev-latest
external
openai/gpt-5.6-luna
typesafe/jev-latest
Registered Aliases
Alias
Model
haiku
anthropic/claude-haiku-4-5
sonnet
anthropic/claude-sonnet-4-6
gpt-mini
openai/gpt-5.6-luna
gpt
openai/gpt-5.6-terra
Decision Models
A decision model is a separate model role from the chat model. It takes a fixed set of facts and a rubric and returns a typed answer — a choice, a score, or a yes/no with a probability — instead of free text. That makes it fast, cheap and predictable for bounded judgments like “is this safe to do without asking?”, and the caller decides what to do with the probability.
Octo’s decision models are served by TypeSafe AI’s hosted Jev models.
Agent
Decision model
hassonic
typesafe/jev-latest
external
typesafe/jev-latest
Where they’re used
Where
Question asked
What happens with the answer
Hass Guard, before each Home Assistant action by a guarded agent (Hassonic or the Octo MCP’s external agent)
Is this action routine and safe to do immediately without asking? (yes/no)
Runs at 85% or higher; otherwise the action is blocked and the agent explains why
The decision_evaluate tool, which lets an agent ask a decision model directly, is not given to any agent. OpenClaw’s experimental decision assistance, which uses a decision model to drop unneeded tools before each turn, is off.
Voice
Settings for speaking replies aloud (text-to-speech) and real-time voice conversations.
Octo Docs