level-1-essentials/individual/05-model-routing/README.md

05 — Model Routing

Which model to use when. The single biggest power-user skill after prompting.

Files in this folder

1. model-routing-cheat-sheet.md — the one-page you print and keep near your monitor
2. provider-strengths-matrix.md — the deeper table with reasoning

The principle

Every model has a shape. Use the right shape for the job.

• Long docs and careful reasoning → Claude
• Real-time facts and citations → Perplexity
• Ecosystem breadth and multimodal → ChatGPT
• Google Workspace and huge context → Gemini
• Cheapest good-enough → Groq (open-weight models)
• Real-time X data or spicy outputs → Grok

Definition of done

• Cheat sheet printed or in a note app you check
• You've mentally sorted your top 5 daily tasks by which model handles each

level-1-essentials/individual/05-model-routing/model-routing-cheat-sheet.md

Model Routing Cheat Sheet

Print this. Tape it above your monitor.

Route by task

Route by data class

Fallback flow

If your first-choice model is rate-limited, down, or the output is bad:

1. Try the second-choice model with the same prompt
2. If both fail, simplify the prompt (fewer variables, smaller scope)
3. If it's a research question, try Perplexity for the base facts, then hand to Claude for synthesis
4. If it's a code question, add the exact error message to the prompt and try Claude Opus

Cost mental model

Approximate cost per million tokens output (as of August 2026 — verify current pricing):

Rule: Start with workhorse. Escalate to frontier only when workhorse can't do it.

level-1-essentials/individual/05-model-routing/provider-strengths-matrix.md

Provider Strengths Matrix

The deep version. Read once, refer back quarterly.

OpenAI (ChatGPT + API)

Signature strengths - Widest ecosystem: Custom GPTs, marketplace, integrations - Best multimodal breadth: image gen (DALL·E / Sora Images), video (Sora), voice (advanced voice mode), Whisper - Structured output: JSON mode is the most reliable - o-series (o1, o3, o4, o5-preview) for reasoning-heavy tasks - Sora for video generation - Massive third-party integration surface

Signature weaknesses - Less nuanced than Claude for long, subtle prose - Custom GPT memory is opaque compared to Claude Projects - Rate limits on peak models can bite

Best for - Multimodal work (image + text) - Custom GPTs shared with others - Batch structured extraction (JSON mode) - Voice interactions - When you need "just works" across every modality

Anthropic Claude

Signature strengths - Best long-form reasoning and prose - Longest sustained coherence over 100k+ token inputs - Artifacts (inline runnable code / documents) - Projects (persistent knowledge + custom instructions) - Claude Code (best agentic coding tool) - Skills (reusable capability packages) - Extended thinking mode surfaces reasoning

Signature weaknesses - No native image generation (as of Aug 2026) - Smaller integration ecosystem than OpenAI - Fewer built-in tools

Best for - Long documents (contracts, reports, research) - Complex coding (via Claude Code) - Writing that needs voice preservation - Anything requiring subtle judgment - Skills-based workflows

Google Gemini

Signature strengths - Longest context (2M tokens on Pro tiers) - Native Google Workspace integration (Docs, Sheets, Gmail, Meet) - Deep Research mode - Veo 3 for video - Multimodal — best at video input - Free tier is generous

Signature weaknesses - Product surface changes often (feature churn) - Consumer UX inconsistent - Voice mode less polished than OpenAI's

Best for - Workspace-native workflows (Google users) - Very long context needs (2M tokens) - Video understanding (feed a video, ask questions) - Research via Deep Research

Perplexity

Signature strengths - Search-grounded by default — every answer cites sources - Real-time facts - Comet browser + Computer (this product) — agentic web workflows - Spaces (persistent contexts with connectors) - Deep Research - API returns citations natively

Signature weaknesses - Less "raw" LLM power — it's a search-tuned system - Long-doc reasoning is not its strength - Some models better for pure generation tasks

Best for - Any question about current events, prices, people - Research briefs with citations - Agent workflows requiring web browsing - Fact-checking

Microsoft Copilot

Signature strengths - Native in Word, Excel, Outlook, Teams, Windows - Best-in-class for Office-native workflows - SharePoint / OneDrive grounding - Enterprise controls and compliance

Signature weaknesses - Feels bolted-on outside Office - Less flexible for creative or open-ended work - Locked to Microsoft's model choices

Best for - Excel formulas, PowerPoint decks, Word docs, Outlook triage - Enterprises already on M365 - Compliance-sensitive environments

xAI Grok

Signature strengths - Real-time X (Twitter) data - More permissive on edgy topics - Aurora image gen without heavy filters

Signature weaknesses - Smaller ecosystem - Fewer integrations - Less careful reasoning than Claude/GPT

Best for - Real-time X monitoring - Uncensored image gen (with judgment) - Contrarian second opinion

Groq (inference provider)

Signature strengths - Fastest inference on the market - Open-weight models (LLaMA, Mixtral, Qwen) - Cheap - Great free tier

Signature weaknesses - Only serves open models — no Claude, GPT, Gemini - Model catalog rotates

Best for - Batch processing where speed matters - Cost-sensitive workloads - When latency is critical (real-time apps)

OpenRouter

Signature strengths - One API key → 300+ models across all providers - Easy A/B testing - Automatic fallback across providers - Great for experimentation

Signature weaknesses - Adds a small margin over provider prices - Rate limits inherited from underlying provider

Best for - Experimentation - Multi-model apps - Fallback resilience

Local models (via Ollama, LM Studio, MLX)

Signature strengths - Zero data leakage - No per-token cost - Works offline - Available: Llama 3.x, Qwen 3, Mistral, DeepSeek, Phi

Signature weaknesses - Much weaker than frontier models - Requires decent hardware (M-series Mac or good GPU) - No multimodal parity

Best for - Highly sensitive data - Offline scenarios - Learning how models work under the hood