Models

Caveman Code runs against any model your provider exposes. The defaults are chosen per provider and re-evaluated on each release based on the proof-bench eval harness.

Default models per provider

Provider	Default model	Editor model (architect mode)
Anthropic	claude-sonnet-4	claude-haiku-4
OpenAI	gpt-5-codex	gpt-5-mini
Google	gemini-2.5-pro	gemini-2.5-flash
OpenRouter	anthropic/claude-sonnet-4	varies
Groq	llama-3.3-70b-versatile	llama-3.1-8b-instant
Cerebras	qwen-2.5-coder-32b	llama-3.1-8b

Override per session:

caveman --model claude-opus-4-7
caveman --model openai/gpt-5
caveman --model claude-sonnet-4:high   # thinking level high

Inside the TUI, /model opens the picker.

Thinking levels

Models that support extended thinking accept a suffix:

Level	Use case
:off	Fastest, lowest cost. Default.
:minimal	Light reasoning. Routine edits.
:low	Default thinking budget for most providers.
:medium	Multi-file refactors.
:high	Cross-cutting concerns, architectural changes.
:xhigh	Hard debugging, complex algorithms.

Cycle in TUI with Shift+Tab.

Architect / editor split

Use a strong model to plan, a cheaper model to execute. Drops cost ~3-5× on long sessions.

/architect set architectModel=claude-opus-4-7 editorModel=claude-haiku-4

Or in ~/.cave/settings.json:

{
    "model": "claude-sonnet-4",
    "modes": {
        "architect": { "model": "claude-opus-4-7" },
        "editor": { "model": "claude-haiku-4" }
    }
}

Per-subagent models

Each subagent declares its own model:

---
description: "Run unit tests and report failures"
model: "claude-haiku-4"   # cheap, since it just shells out
tools: [Bash, Read]
---

Subagent results are summarized to ≤500 tokens before re-entering the parent's context — letting you spend on Haiku instead of Opus for repetitive subtasks.

Model registry

Provider/model definitions live in a versioned registry at github.com/cave-cli/registry. Update locally without releasing Caveman Code:

caveman models update

Override per-machine in ~/.cave/registry.json. See Provider Registry for schema.

Cost-aware defaults

Caveman Code's first-run wizard suggests Haiku/Flash for the default model on free OAuth accounts to avoid surprise bills. Upgrade with /model once you've validated the workflow.