Model selection
Choose a model for Claude, OpenCode, Codex, Copilot, or Cursor.
Claude models
| Model | When to pick it |
|---|---|
| Sonnet 5.5 | Balanced cost and quality, faster and cheaper than Sonnet 4.6. |
| Opus 5.5 | Complex APIs (many endpoints, nested auth, polymorphic responses). |
| Opus 4.6 | The previous Opus, for comparison or when a run needs it. |
| Haiku 5.5 | Cheapest and fastest. Good for simple endpoints or quick iterations. |
Sonnet 4.6 stays the default until you choose another model. It and Haiku 4.5
are no longer offered in the pickers, but --model and config.json still
accept them.
Change the model
In the session, type /model.
Or pass a model to a direct command:
reverse-api-engineer agent --prompt "capture the jobs api" --model claude-sonnet-5-5Or edit ~/.reverse-api/config.json directly:
{
"claude_code_model": "claude-opus-5-5"
}Effort
Claude models take an effort level, from low to max: how hard the agent
thinks before acting. Higher effort is slower and costs more, and helps on hard
captures. Set it with /effort or Ctrl+E in the session, or with
claude_effort in the config. Leaving it unset keeps each model's default:
medium for Opus 5.5 and Haiku 5.5, high for the others. Haiku 4.5 has no
effort setting.
Config file reference
The full default config:
{
"agent_provider": "chrome-mcp",
"agent_browser_notes": "",
"agent_browser_npx_package": "agent-browser@0",
"claude_code_model": "claude-sonnet-4-6",
"claude_setting_sources": [],
"claude_effort": null,
"cursor_model": "composer-2.5",
"cursor_web_search": true,
"cursor_setting_sources": null,
"copilot_model": "gpt-5",
"codex_model": null,
"codex_effort": null,
"codex_home": null,
"opencode_model": "big-pickle",
"opencode_provider": "opencode",
"opencode_auto_start": true,
"opencode_base_url": "http://127.0.0.1:4096",
"opencode_npx_package": "opencode-ai@latest",
"ollama_auto_start": true,
"ollama_base_url": "http://127.0.0.1:11434",
"output_dir": null,
"output_language": "python",
"real_time_sync": true,
"sdk": "claude"
}OpenCode users
If you've configured the OpenCode SDK, choose opencode provider / model in
/settings. RAE loads the connected providers and active tool-capable models
from your OpenCode server, marks free OpenCode options, and saves the pair
together. Advanced users can still edit the opencode_model and
opencode_provider keys directly.
For Ollama, set opencode_provider to ollama. The /settings flow discovers
installed models and only offers those that report tool calling plus the 64k
context OpenCode requires. RAE does not pull models automatically.
Codex users
Codex uses codex_model. Its default, null, runs Codex's default model for
your account. Choose codex model in /settings to pick from the models
your account offers, or set a model id such as "gpt-6.1-sol" in the config.
Codex models take their own effort levels, which differ from model to model.
/effort in a Codex session, or codex effort in /settings, lists the
levels the model offers; codex_effort: null keeps the model's default.
Changing the model drops an effort the new model does not offer.
Copilot and Cursor users
Copilot uses copilot_model (default gpt-5). Cursor uses cursor_model
(default composer-2.5) plus cursor_web_search / cursor_setting_sources
for controlling which Cursor settings layers are loaded.