Models

TatsuCode lets you switch between cloud, subscription, and local models quickly.

Use:

/models

This opens the model selector with filtering and capability labels.

Provider Groups

Models are grouped by source:

• OpenAI Plus/Pro (subscription OAuth)
• Claude Code (Claude subscription through Anthropic's claude CLI)
• OpenRouter (API key)
• Local Providers (llama.cpp, Ollama, LM Studio, custom endpoints)

Decision models (TypeSafe AI) are a separate kind of model: they answer typed questions rather than chat, so they do not appear in the picker. The agent uses them through the TypeSafe tool.

Some provider routes can be temporarily limited by upstream authorization status. If one route is unavailable, use another provider (for example, Gemini via OpenRouter).

Notable Model Families

OpenAI Plus / Pro (direct)

Reasoning effort goes up to Max on every model in this list.

GPT-5.5, GPT-5.4 (including Mini and Nano), and GPT-5.3 Codex (including Spark) have been retired from Codex. If one of them was your selected model or a Task Agent model, TatsuCode moves you to GPT-6.1 Sol, or to GPT-6 Luna for the small models.

Claude Code (subscription)

These run the full Claude Code agent (with Claude's own native tools) through Claude Agent using your Pro / Max / Team / Enterprise subscription. Depending on your plan, Fable-tier models may draw on pay-as-you-go usage credits; Anthropic bills you directly, as always. Sonnet 5.5 needs Claude CLI v2.1.284 or later and Haiku 5.5 needs v2.1.293 or later; run claude update if they fail to start.

OpenRouter (curated defaults)

OpenRouter exposes hundreds of models; TatsuCode curates a default lineup of the ones that work reliably with its tool calling and streaming, and auto-includes new releases that meet the same bar. /models-add shows the full catalog.

Local

Any GGUF or OpenAI-compatible server works. Popular picks: Gemma 4, Qwen 3.5, Kimi K2.5, DeepSeek. Multimodal variants (Gemma 4, Qwen-VL) are detected automatically for vision.

Model Capabilities

Context Window

The context window controls how much code and history the model can process in one turn.

• ~1M context: every Claude Code model (Haiku included), GPT-5.4 via OpenRouter, Gemini 3.1 Pro. Excellent for wide codebase understanding
• 872K context: every OpenAI Plus/Pro model (the GPT-6 and GPT-5.6 families)
• 272K to 400K context: most OpenRouter GPT-5.x models. Great for serious multi-file coding tasks
• 200K context: Claude Haiku 4.5 and most OpenRouter Claude models. Strong for most production workflows

Vision

Many flagship models support image input for:

• UI screenshot debugging
• visual regression checks
• diagram or mockup understanding

Image detail modes on OpenAI Direct GPT-5.4 and later (/image-detail) let you send screenshots at up to their original resolution, which matters for dense diagrams and high-DPI mockups.

Local-provider vision auto-detection — local models that support vision (Gemma 4, Qwen-VL, and other multimodal variants) are detected automatically by querying the local server's metadata, with self-healing retry if the server starts after TatsuCode. No manual capability flag required.

Image Generation Models

A separate model lineup is used for image generation, configured via /image-generation:

See Images for the full workflow.

Reasoning Controls

For supported models, tune reasoning depth:

/reason-effort
/reason-display

Use higher effort for complex debugging and planning; lower effort for speed.

Choosing the Right Model

Quick Pick Guide

Cost Strategy

• Use your connected subscription routes first
• Keep OpenRouter for breadth and fallback
• Use local models for private/offline workflows
• Assign cheaper models to scout-style sub-agents via /models-taskagents — see Task Agents

OpenRouter list curation: the default OpenRouter lineup is trimmed to models that perform well with TatsuCode's tool calling and streaming. Anything else in OpenRouter's catalog is one /models-add away.

Model List Management

/models-add
/models-remove

• Add only what you actively use
• Keep a "fast + deep" pair in your favorites
• Switch models by task rather than sticking to one for everything

Usage and Quota

/usage
/usage-quota

• /usage shows session token usage
• /usage-quota shows subscription/API limit information where supported

If context grows too large:

/compact

Recent builds improved auto-compaction reliability and long-session recovery.

Next Steps

• Providers — connect and manage provider access
• Settings — reasoning, temperature, and display options
• AI Capabilities — what TatsuCode can do with these models