Docs
← Back to site Download

AI Models

VibeRails supports multiple AI models across two CLI backends. Choose the right model for your review based on depth, speed, and cost requirements.

Overview

VibeRails does not call AI APIs directly. Instead, it orchestrates local CLI tools — Claude Code and Codex CLI — which handle authentication and model access. You select a CLI backend and model in the review wizard (step 2) or fix session wizard (step 3).

AI model selection cell in the reviews table
Each review shows which CLI and model were used, visible in the reviews table.

Claude Code Models

When you select Claude Code as the CLI backend, the following models are available:

Model Strengths Speed Extended Thinking
Opus 4.5 Deepest analysis, best at finding subtle bugs and security issues. Strongest reasoning across all categories. Slowest Available
Sonnet 4 Excellent balance of quality and speed. The recommended default for most reviews. Fast Available
Haiku Fastest and most cost-effective. Good for quick scans and initial passes on large codebases. Fastest Not available

Extended Thinking

Opus 4.5 and Sonnet 4 support Extended Thinking, which gives the model additional time to reason through complex problems before responding. When enabled, the model produces more thorough analysis at the cost of longer review times.

Extended Thinking is especially valuable for:

  • Security audits where subtle vulnerabilities require multi-step reasoning
  • Complex concurrency or state-management code
  • Large files with interrelated logic across many functions
Tip: Start with Sonnet 4 (no extended thinking) for everyday reviews. Switch to Opus 4.5 with Extended Thinking for security-critical or pre-release audits.

Codex CLI Models

When you select Codex CLI as the backend, the following OpenAI models are available:

Model Strengths Speed Reasoning Level
O3 Most capable OpenAI reasoning model. Excels at complex logic, mathematical reasoning, and nuanced code analysis. Slowest Low / Medium / High
O3-mini Balanced reasoning model. Good quality with faster throughput than full O3. Medium Low / Medium / High
O4-mini Latest compact reasoning model. Fast and cost-effective with strong code understanding. Fast Low / Medium / High

Reasoning Levels

All Codex CLI models support configurable Reasoning Levels that control how much compute the model spends thinking before responding:

Level Behaviour Best For
Low Minimal reasoning. Fast responses with basic analysis. Quick scans, style checks, obvious bugs.
Medium Moderate reasoning. Balanced depth and speed. General-purpose reviews. The recommended default.
High Maximum reasoning. Deepest analysis with longest response time. Security audits, complex algorithm review, critical-path code.
Tip: Reasoning Level "Medium" on O4-mini is a great default for Codex users — it offers strong results at good speed and cost.

When to Use Which Model

The right model depends on your priorities. Here is a decision guide:

Scenario Recommended Model Settings
Everyday code review Sonnet 4 No extended thinking
Quick scan of a large codebase Haiku or O4-mini Low reasoning
Security audit Opus 4.5 or O3 Extended thinking / High reasoning
Pre-release review Sonnet 4 or O3-mini Extended thinking / Medium reasoning
Fix session (generating patches) Sonnet 4 or O4-mini Default settings
Cost-sensitive / high-volume Haiku or O4-mini Low reasoning

Configuration in Wizards

You select the AI model at two points in the VibeRails workflow:

Review Wizard (Step 2)

When creating a new code review, step 2 of the wizard lets you pick the CLI backend, model, and thinking/reasoning options. Your selection is saved with the review so you can compare results across different models.

Fix Session Wizard (Step 3)

Fix sessions also let you choose a model. You may want a different model for fixes than for reviews — for example, using Sonnet 4 for review but Haiku for generating quick patches.

Tip: You can set default models for both reviews and fixes in Settings, so you only need to override in the wizard when you want something different.