Models

Agentastic supports multiple AI models, each optimized for different tasks. Understanding which model to use and when will help you get the best performance and results.

Model Types

Agentastic categorizes models into six capability types, each optimized for different tasks:

Chat Models 💬

General-purpose conversational models for everyday tasks:

• Best for: Quick queries, text generation, casual conversations
• Speed: Fast response times
• Tools: Limited tool access
• Examples: GPT-4o, Claude Haiku, Agentastic Chat

When to use: Default choice for most interactions, simple text expansions, quick questions.

Vision Models 👁

Models that can understand and analyze images:

• Best for: Screenshot analysis, image descriptions, visual Q&A
• Speed: Moderate response times
• Tools: Standard tools plus image processing
• Examples: GPT-4o, Claude Sonnet, Qwen3 VL, Agentastic Vision

When to use: Analyzing screenshots, understanding diagrams, extracting text from images.

Agent Models ∞

Advanced models with reasoning and tool-use capabilities:

• Best for: Complex tasks, multi-step workflows, tool integration
• Speed: Slower but more capable
• Tools: Full access to all tools (calendar, email, search, files)
• Examples: Agentastic Agent, Claude Opus, GPT-5, Qwen3 Max

When to use: Complex research, calendar management, email composition, file operations.

Automation Models 🖱️

Specialized models for desktop control and automation:

• Best for: GUI automation, desktop control, system tasks
• Speed: Variable depending on task
• Tools: Desktop control, mouse/keyboard automation
• Examples: Agentastic Computer Use, Claude Sonnet, OpenAI Computer Use

When to use: Automating repetitive desktop tasks, UI testing, complex workflows.

Thinking Models 🧠

Models with advanced reasoning and "thinking" capabilities:

• Best for: Deep reasoning, complex problem-solving, multi-step analysis
• Speed: Slower due to extended reasoning
• Tools: Full tool access with enhanced reasoning
• Examples: OpenAI O3, DeepSeek R1, Qwen3 Thinking, Claude Opus

When to use: Complex problem-solving, research requiring deep reasoning, multi-step analysis.

Image Generation Models 🎨

Models that can generate and edit images:

• Best for: Creating images from text, image editing
• Speed: Variable depending on complexity
• Tools: Image generation and editing
• Examples: Agentastic Imagine, Gemini Banana, Hunyuan v3, SeeDream v4

When to use: Creating images from descriptions, editing existing images, visual content generation.

Available Models

Agentastic Recommended Models 💜

These are Agentastic's optimized models, recommended for most users:

OpenAI Models

Anthropic (Claude) Models

Google Models

Gemini 3 update (Nov 18, 2025)

• Launch blog highlights that Gemini 3 Pro now tops major benchmarks (1501 Elo on LMArena, 81% on MMMU-Pro, 87.6% on Video-MMMU, 72.1% on SimpleQA Verified) and introduces a Deep Think mode that reaches 41% on Humanity's Last Exam (no tools) and 45.1% on ARC-AGI-2 with code execution. Deep Think is rolling out to Gemini Ultra subscribers after safety testing.
• The Gemini 3 API guide documents the gemini-3-pro-preview model ID, a 1M input / 64K output window, a Jan 2025 knowledge cutoff, and pricing of $2 / $12 per 1M tokens below 200K tokens ($4 / $18 after). It also introduces the thinking_level parameter (low for lower latency/cost, high default for Deep Think depth—medium is coming soon) and the new media_resolution_{low|medium|high} knobs: images default to 1120 tokens at high, PDFs saturate at 560 tokens with medium, and video frames cost 70 tokens at low/medium (280 tokens if you force high for dense OCR).
• Gemini 3 Pro does not yet support pixel-level segmentation; keep /model:flash (Gemini 2.5 Flash with thinking disabled) or Gemini Robotics-ER 1.5 for those workloads per Google's guidance.

DeepSeek Models

Grok Models (x-ai)

Qwen Models

Meta (Llama) Models

Other Providers

Image Generation Models

Model Selection

Quick Selection in Launcher

Use the /model: tag to quickly switch models:

1. Type /model: in the launcher
2. Browse available models with arrow keys
3. Select with Enter or click
4. Continue with your prompt

Example:

/model:opus Analyze this research paper and summarize key findings
/model:agent What's on my calendar today?
/model:vision What do you see in this screenshot?

Setting Default Models

Configure your preferred models in Settings:

1. Open Settings (⌘,)
2. Go to AI Models tab
3. Set defaults for each model type:
   • Default Chat Model
   • Default Vision Model
   • Default Agent Model
   • Default Automation Model
   • Default Thinking Model
   • Default Image Generation Model

Model-Specific Tags

Some tags automatically suggest appropriate models:

• /vision - Activates vision-capable models
• /agent - Enables agent mode with tool access
• /automation - Uses automation-capable models
• /imagine - Uses image generation models

Configuration

Models Configuration File

Advanced users can customize models in ~/Library/Application Support/agentastic/config/models.yml:

active_models:
  - value: agentastic:agent
    label: Agentastic Agent 💜
    description: Designed for advanced agentic tasks
    context_length: 256000
    capabilities:
      - agent
      - chat
    fallback_model: agentastic:sonnet

  - value: agentastic:chat
    label: Agentastic Chat 💜
    description: The friendly neighborhood model
    context_length: 128000
    capabilities:
      - chat
    fallback_model: agentastic:sonnet

  - value: agentastic:vision
    label: Agentastic Vision 💜
    description: Optimized vision analysis model
    context_length: 200000
    capabilities:
      - chat
      - agent
      - vision
    fallback_model: agentastic:sonnet

API Keys

Configure API keys for different providers:

1. Open Settings → AI Models
2. Click Configure API Keys
3. Enter keys for:
   • OpenAI (for GPT models)
   • Anthropic (for Claude models)
   • Google (for Gemini models)
   • DeepSeek, Grok, Qwen, and other providers (if configured)

Security Note: API keys are stored securely in your macOS keychain.

Model Selection Strategy

For Different Tasks

Writing Emails

• Quick email: Agentastic Chat, GPT-4o, Claude Haiku
• Important email: Agentastic Writer, Claude Sonnet, GPT-5
• Complex negotiation: Claude Opus, Qwen3 Max

Research & Analysis

• Quick fact check: Agentastic Auto, GPT-4o
• Deep research: Claude Opus, Qwen3 Max with /agent
• Visual analysis: Agentastic Vision, GPT-4o, Qwen3 VL

Code & Technical

• Quick snippets: Agentastic Chat, GPT-4o
• Complex debugging: Qwen3 Coder, Grok Code Fast, Claude Opus
• Code review: Claude Sonnet, Agentastic Writer

Automation

• Simple tasks: Agent models with tools
• Desktop control: Agentastic Computer Use, OpenAI Computer Use
• Multi-step workflows: Agentastic Agent, Claude Opus with /agent

Performance vs. Quality

Balance speed and capability based on your needs:

Fast Response Needed → Agentastic Chat, GPT-4o, Claude Haiku, GPT-4o-mini

Balanced Performance → Agentastic Auto, Claude Sonnet, GPT-5, Gemini 3 Pro (Preview)

Best Quality → Claude Opus, GPT-5, Qwen3 Max, OpenAI O3

Special Capabilities → Vision models for images (Agentastic Vision, Qwen3 VL), Automation models for desktop control (Agentastic Computer Use), Thinking models for deep reasoning (O3, DeepSeek R1), Image generation (Agentastic Imagine, Gemini Banana)

Cost Considerations

Different models have different pricing:

Cost Tiers (Relative)

1. Low Cost: Agentastic Chat, Claude Haiku, GPT-4o-mini (for quick tasks)
2. Medium Cost: Agentastic Auto, Claude Sonnet, GPT-4o, Gemini 2.5 Flash
3. Higher Cost: Claude Opus, GPT-5, Gemini 3 Pro (Preview), Qwen3 Max, Agentastic Agent
4. Specialized: Automation models, Thinking models, Image generation models (variable)

Optimizing Costs

1. Use appropriate models: Don't use Opus for simple tasks
2. Leverage chat mode: Agent mode uses more tokens
3. Clear context: Start new conversations when context isn't needed
4. Local models: Consider Ollama for high-volume use

Model Limitations

Context Windows

Different models have different context limits:

• Small (8K-65K): OpenAI Computer Use (8K), GLM Vision (65K)
• Standard (128K-131K): Most models including GPT-5, GPT-4o, DeepSeek
• Large (200K-256K): Claude models (200K), Agentastic Agent (256K), Grok models (256K), Qwen3 Max (256K)
• Very Large (262K): Qwen3 Thinking, Qwen3 Next models
• Massive (1M+): Google Gemini 3 Pro (Preview) / Gemini 2.5 Flash (1M tokens), Grok 4 Fast (2M tokens)

Rate Limits

API providers impose rate limits:

• Per minute: Number of requests
• Per day: Total token usage
• Concurrent: Parallel requests

Agentastic automatically manages rate limits and queues requests when necessary.

Feature Support

Not all models support all features:

*Depends on specific model

Troubleshooting

Model Not Available

If a model isn't showing up:

1. Check API key is configured
2. Verify internet connection
3. Check provider service status
4. Ensure model is enabled in settings

Poor Results

If a model isn't performing well:

1. Try a more capable model
2. Use appropriate tags (/agent, /vision)
3. Provide clearer prompts
4. Add relevant context via memory channels

Rate Limit Errors

If you hit rate limits:

1. Wait for the reset period (usually 1 minute)
2. Switch to a different model
3. Reduce request frequency
4. Consider upgrading API tier

Best Practices

1. Start with Agentastic Auto or recommended models (💜) for most tasks
2. Use vision models only when analyzing images
3. Enable agent mode when you need tools
4. Reserve automation models for desktop tasks
5. Use thinking models for complex reasoning problems
6. Use image generation models for creating visual content
7. Test different models to find what works best for your workflow
8. Monitor usage to optimize costs

What's Next?

Now that you understand models:

• Core Features - Explore all capabilities
• AI Commands - Master special tags
• Advanced Settings - Fine-tune configuration
• Troubleshooting - Solve common issues

Pro Tip: Use /model: frequently to experiment with different models. You'll quickly learn which models work best for your specific use cases.