Model selection

Model selection determines everything: cost, performance, privacy, and capabilities. With 50+ providers and 500+ models, choosing the right one feels overwhelming. This guide cuts through the noise.

Selection Framework

Choose models based on three criteria:

Task — What are you actually doing? (coding, writing, analysis, reasoning)
Context — How much text do you need to process? (10K vs 200K vs 1M tokens)
Privacy — Is data sensitive? (local vs cloud)

Model Categories

Category	Best For	Examples
Coding	Code generation, debugging, refactoring	Claude Sonnet 4, GPT-5.4, CodeLlama
Reasoning	Math, logic, planning	Gemini 3.1 Pro, GPT-5.4, o1
Vision	Image understanding, OCR	Gemini 3.1 Pro Vision, Claude 3.5 Sonnet
Writing	Content, copy, editing	GPT-5.3, Claude Opus 4.7 4, Claude Sonnet 4
Chinese	Chinese language tasks	GLM 5, Kimi 2.5, DeepSeek
Local	Privacy, offline, zero cost	Ollama (LLaMA 3.2), LM Studio

Context Window Guide

10K tokens: Short messages, quick queries, simple tasks
200K tokens: Full codebases, long conversations, documents
1M tokens: Entire codebases, multi-file generation, vision + text

Cost Comparison

Provider	Model	Cost/1M tokens
Anthropic	Claude Sonnet 4	$3-5
OpenAI	GPT-5.4	$5-8
Google	Gemini 3.1 Pro	$2-4
OpenRouter	Free tier	$0
Ollama	LLaMA 3.2	$0 (local)

💡 Free tier — OpenRouter offers Grok 4.20:free, Gemini 3.1 Pro Preview:free, DeepSeek R1:free. Real models, not trials.

Selection Examples

# Coding task
Claude Sonnet 4 — Best for full IDE integration
GPT-5.4 — Good for general coding
CodeLlama — Local, free

# Reasoning task
Gemini 3.1 Pro — 1M context, strong reasoning
GPT-5.4 — General reasoning
o1 — Specialized reasoning

# Vision task
Gemini 3.1 Pro Vision — 1M context, vision
Claude 3.5 Sonnet — Vision + coding

# Privacy-sensitive
Ollama LLaMA 3.2 — 100% local
LM Studio — Local GUI

Testing Strategy

Before committing to a paid model, test with free tiers:

Start with OpenRouter free tier
Test on your actual use case
Compare quality vs cost
Switch to paid if it saves time

The Bottom Line

Start with the free tier of OpenRouter or local models. Only pay when you hit limits. Match the model to the task — don't use a sledgehammer for a screw.