Skip to content

AI Models

Use the right tool for each job—fast models for quick tasks, smart ones for hard problems.

Most AI tools lock you into one model. Fabric lets you choose. Running a quick lint check? Use a fast, cheap model. Debugging a gnarly race condition? Switch to something powerful. You control the cost-quality tradeoff.

No More Overpaying

Stop using expensive models for simple tasks. Fabric's model selector puts you in control of your AI costs.

Model Tier System

Fabric automatically selects models based on task complexity using a three-tier system:

Tier Selection

Models are ranked by a formula that balances intelligence (SWE-bench score), speed (tokens/sec), and efficiency (latency and cost):

Score = SWE-bench × (tokens/sec) / (latency × totalCost)

Large Tier - Most Capable

Used for complex reasoning, detailed planning, and architecture decisions.

Trade-off: Intelligence > Speed/Cost

Model Provider SWE-bench Cost/M Tokens
claude-opus-4-5 Anthropic 72 $30
gpt-5-pro OpenAI 70 $135
claude-sonnet-4-5 Anthropic 68 $18
gemini-3-pro-preview Google 65 $14
gpt-5.1 OpenAI 62 $11.25
gpt-5.1-codex OpenAI 60 $11.25
deepseek-reasoner Deepseek 55 $2.74
qwen3-235b-thinking OpenRouter 52 $0.71

Medium Tier - Balanced

Used for targeted code changes, merging, and code review.

Trade-off: Balance of speed, cost, and capability

Model Provider SWE-bench Tokens/sec Cost/M Tokens
claude-haiku-4-5 Anthropic 48 200 $6
gpt-5.1-codex-mini OpenAI 45 150 $2.25
zai-glm-4.6 Cerebras 42 2000 $5
qwen-3-235b-instruct Cerebras 40 1500 $1.80
deepseek-chat Deepseek 38 100 $1.37
qwen3-coder-30b OpenRouter 35 120 $0.31
grok-code-fast-1 OpenRouter 33 300 $1.70

Small Tier - Fast & Cheap

Used for summaries, tab naming, and high-volume background operations.

Trade-off: Speed/Cost > Intelligence

Model Provider SWE-bench Tokens/sec Cost/M Tokens
qwen-3-32b Cerebras 30 2500 $1.20
gpt-oss-120b Cerebras 28 2200 $1.10
gemini-2.5-flash Google 32 500 $0.75
gpt-5-nano OpenAI 25 300 $0.45
deepseek-r1:free OpenRouter 35 80 Free
deepseek-v3.2-speciale OpenRouter 55 100 $0.68
local models Local 20 50 Free

Supported Providers

Anthropic

Claude models are known for thoughtful, nuanced responses and excellent code generation.

Model Display Name Context Output Vision Reasoning Intelligence
claude-opus-4-5 opus 4.5 200k 32k On/Off ⭐⭐⭐⭐⭐
claude-sonnet-4-5 sonnet 4.5 200k 64k On/Off ⭐⭐⭐⭐
claude-haiku-4-5 haiku 4.5 200k 64k On/Off ⭐⭐⭐

Capabilities

  • Vision support (images, PDFs)
  • Extended thinking mode
  • Tool use
  • JSON mode
  • Web search
  • :material-stream: Streaming responses

:simple-openai: OpenAI

GPT-5 series models with advanced reasoning capabilities.

Model Display Name Context Output Vision Reasoning Intelligence
gpt-5-pro gpt 5 pro 400k 128k Effort Levels ⭐⭐⭐⭐⭐
gpt-5.1 gpt 5.1 400k 128k Effort Levels ⭐⭐⭐⭐
gpt-5.1-codex gpt 5.1 codex 400k 128k Effort Levels ⭐⭐⭐⭐
gpt-5.1-codex-mini gpt 5.1 codex mini 400k 128k Effort Levels ⭐⭐⭐
gpt-5-nano gpt 5 nano 400k 128k ⭐⭐

Reasoning Effort Levels

OpenAI models support configurable reasoning effort: low, medium, high

Intelligence ratings vary by effort level (e.g., gpt-5-pro: high=5★, medium=5★, low=4★)

Capabilities

  • Vision support (images, PDFs)
  • Extended thinking with effort control
  • Tool use
  • JSON mode
  • Web search
  • :material-stream: Streaming responses

Google

Gemini models with massive context windows and code execution.

Model Display Name Context Output Vision Reasoning Intelligence
gemini-3-pro-preview gemini 3.0 pro 1M 64k On/Off ⭐⭐⭐⭐⭐
gemini-2.5-flash gemini 2.5 flash 1M 65k ⭐⭐⭐

Capabilities

  • Vision support (images, PDFs)
  • Extended thinking mode (gemini-3-pro only)
  • Tool use
  • JSON mode
  • Web search
  • Code execution (unique to Google)
  • :material-stream: Streaming responses

:material-brain-circuit: Cerebras

Ultra-fast inference models optimized for code generation.

Model Display Name Context Output Reasoning Intelligence
zai-glm-4.6 ZAI GLM 4.6 (preview) 131k 8k ⭐⭐⭐⭐
gpt-oss-120b GPT-OSS 120B 131k 8k ⭐⭐⭐
qwen-3-32b Qwen 3 32B 131k 8k On/Off ⭐⭐⭐
qwen-3-235b-instruct Qwen 3 235B Instruct (preview) 131k 8k On/Off ⭐⭐⭐⭐

Performance

Cerebras models deliver 2000-2500 tokens/sec - the fastest in Fabric's lineup.

Capabilities

  • Tool use
  • :material-stream: Streaming responses
  • Extended thinking (select models)
  • No vision support
  • No PDF support

Deepseek

Specialized reasoning models with excellent code understanding.

Model Display Name Context Output Reasoning Intelligence
deepseek-reasoner deepseek reasoner 128k 64k On/Off ⭐⭐⭐
deepseek-chat deepseek chat 64k 4k On/Off ⭐⭐⭐

Capabilities

  • Extended thinking mode
  • Tool use
  • JSON mode
  • :material-stream: Streaming responses
  • No vision support

OpenRouter

Access to diverse open-source and commercial models through a single API.

Model Display Name Context Output Reasoning Intelligence
x-ai/grok-code-fast-1 Grok Code Fast 1 256k 128k ⭐⭐⭐⭐
deepseek/deepseek-v3.2-speciale DeepSeek V3.2 Speciale 164k 65k On/Off ⭐⭐⭐⭐
qwen/qwen3-235b-thinking Qwen3 235B Thinking 262k 81k On/Off ⭐⭐⭐⭐
z-ai/glm-4.6 GLM 4.6 202k 100k ⭐⭐⭐⭐
minimax/minimax-m2 Minimax M2 204k 100k ⭐⭐⭐
qwen/qwen3-coder-30b Qwen3 Coder 30B 262k 128k ⭐⭐⭐
deepseek/deepseek-r1:free DeepSeek R1 164k 128k On/Off ⭐⭐⭐

Capabilities

  • Tool use
  • :material-stream: Streaming responses
  • Extended thinking (select models)
  • Free tier available (deepseek-r1)

Limitations

  • No vision support (most models)
  • No JSON mode
  • No web search

Local Models

Run models locally using llama.cpp, Ollama, or custom endpoints.

Feature Status
Endpoint http://localhost:8080/v1/
Context Model-dependent
Output Model-dependent
Cost Free
Intelligence ⭐⭐ (typical)

Local Setup

Configure local providers through Settings → Models → Add Provider

Supported backends:

  • llama.cpp - High-performance C++ inference
  • Ollama - Easy model management
  • Custom - Any OpenAI-compatible endpoint

Limitations

  • No vision support
  • No tool use
  • No JSON mode
  • No reasoning mode
  • :material-stream: Streaming supported

Model Capabilities

Vision Support

Models that can process images and PDFs:

  • All Anthropic models (Claude 4.5 series)
  • All OpenAI models (GPT-5 series)
  • All Google models (Gemini)

Supported Formats

  • Images: PNG, JPEG, GIF, WebP
  • Documents: PDF (automatic extraction)

Extended Thinking

Models with chain-of-thought reasoning:

Reasoning Types

  • On/Off - Simple toggle (Anthropic, Google, Deepseek, select others)
  • Effort Levels - Configurable low/medium/high (OpenAI)

Supported models:

  • Anthropic: All Claude 4.5 models
  • OpenAI: gpt-5-pro, gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini
  • Google: gemini-3-pro-preview
  • Cerebras: qwen-3-32b, qwen-3-235b-instruct
  • Deepseek: All models
  • OpenRouter: deepseek-v3.2, qwen3-235b-thinking, deepseek-r1

Tool Use

All models except local models support tool/function calling for:

  • Code execution
  • File operations
  • API integrations
  • Custom tools

Models with integrated web search capabilities:

  • All Anthropic models
  • All OpenAI models
  • All Google models

Code Execution

Only Google Gemini models natively support code execution in the model environment.

Provider Configuration

Adding API Keys

Quick Setup

  1. Go to Settings → Models
  2. Select your provider
  3. Enter your API key
  4. Choose default models for each tier

Custom Endpoints

For self-hosted or custom model endpoints:

Provider: custom
Endpoint: https://your-api.example.com/v1/
API Key: your-key-here

Fabric supports any OpenAI-compatible API endpoint.

Model Selection Strategy

Fabric automatically selects models based on the task:

Task Tier Example Use Case
Planning Large Creating implementation plans, architecture decisions
Code Review Large Reviewing medium model output, complex refactoring
Implementation Medium Writing targeted code changes, merging edits
Quick Tasks Small Summaries, tab naming, background operations

Manual Override

You can manually select any model regardless of tier from the model dropdown.

Cost Optimization

Budget-Friendly Options

Free models:

  • OpenRouter: deepseek-r1:free
  • Local: Any self-hosted model

Low-cost high-performers:

Model Provider Cost/M SWE-bench Speed
qwen3-235b-thinking OpenRouter $0.71 52 Moderate
gemini-2.5-flash Google $0.75 32 Fast
deepseek-chat Deepseek $1.37 38 Moderate
qwen-3-235b-instruct Cerebras $1.80 40 Very Fast

Performance Options

For maximum intelligence regardless of cost:

  1. claude-opus-4-5 - Best reasoning (SWE-bench: 72)
  2. gpt-5-pro - Strong all-around (SWE-bench: 70)
  3. claude-sonnet-4-5 - Excellent balance (SWE-bench: 68)

Frequently Asked Questions

Which model should I use?

For most coding tasks: Start with claude-sonnet-4-5 or gpt-5.1-codex (large tier)

For quick tasks: Use claude-haiku-4-5 or gemini-2.5-flash (small tier)

For complex architecture: Use claude-opus-4-5 or gpt-5-pro (large tier)

On a budget: Use deepseek-r1:free or local models

How does tier selection work?

Fabric ranks models within each tier by:

  1. Intelligence (SWE-bench score)
  2. Speed (tokens per second)
  3. Efficiency (latency and cost)

The highest-ranked available model is auto-selected for each task type.

Can I use multiple providers simultaneously?

Yes! Configure API keys for multiple providers and Fabric will use the best available model for each task based on tier rankings.

Do I need all API keys?

No - configure only the providers you want to use. At minimum, one provider with large, medium, and small tier models is recommended.

What's the difference between reasoning types?
  • On/Off: Simple toggle for extended thinking (Anthropic, Google, Deepseek)
  • Effort Levels: Adjustable low/medium/high reasoning intensity (OpenAI)

Higher effort = better quality but slower responses and higher costs.

Can I add custom models?

Yes! Use the custom provider option with any OpenAI-compatible endpoint. Local models (llama.cpp, Ollama) are also supported.