4.8 KiB
Choosing the Right Model
Different models, different strengths. Here's how to pick.
OpenRouter gives you access to 100+ models. That's overwhelming. Let me cut through the noise.
The Quick Answer
Just getting started? Use these:
| Use Case | Model | Why |
|---|---|---|
| General coding | x-ai/grok-code-fast-1 |
Fast, cheap, capable |
| Complex problems | google/gemini-3-pro-preview |
1M context, solid reasoning |
| Code-specific | openai/gpt-5.1-codex |
Trained specifically for code |
| Budget mode | minimax/minimax-m2 |
Cheapest that actually works |
Pick one. Start working. Switch later if needed.
Discovering Models
Top recommended (curated list):
claudish --top-models
All OpenRouter models (hundreds):
claudish --models
Search for specific models:
claudish --models grok
claudish --models codex
claudish --models gemini
JSON output (for scripts):
claudish --top-models --json
claudish --models --json
Understanding the Columns
When you see the model table:
Model Provider Pricing Context Caps
google/gemini-3-pro-preview Google $7.00/1M 1048K ✓ ✓ ✓
Model - The ID you pass to --model
Provider - Who made it (Google, OpenAI, xAI, etc.)
Pricing - Average cost per 1 million tokens. Input and output prices vary, this is the midpoint.
Context - Maximum tokens the model can handle (input + output combined)
Caps (Capabilities):
- First ✓ = Tools - Can use Claude Code's file/bash tools
- Second ✓ = Reasoning - Extended thinking mode
- Third ✓ = Vision - Can analyze images/screenshots
My Honest Model Breakdown
Grok Code Fast 1 (x-ai/grok-code-fast-1)
Price: $0.85/1M | Context: 256K
My daily driver. Fast responses, good code quality, reasonable price. Handles most tasks without drama.
Good for: General coding, refactoring, quick fixes Bad for: Very long files (256K limit), vision tasks
Gemini 3 Pro (google/gemini-3-pro-preview)
Price: $7.00/1M | Context: 1M (!)
The context king. A million tokens means you can dump entire codebases into context. Reasoning is solid. Vision works.
Good for: Large codebase analysis, complex architecture, image-based tasks Bad for: Quick tasks (overkill), budget-conscious work
GPT-5.1 Codex (openai/gpt-5.1-codex)
Price: $5.63/1M | Context: 400K
OpenAI's coding specialist. Trained specifically for software engineering. Does code review really well.
Good for: Code review, debugging, complex refactoring Bad for: General chat (waste of a specialist)
MiniMax M2 (minimax/minimax-m2)
Price: $0.60/1M | Context: 204K
The budget champion. Cheapest model that doesn't suck. Surprisingly capable for simple tasks.
Good for: Quick fixes, simple generation, high-volume tasks Bad for: Complex reasoning, architecture decisions
GLM 4.6 (z-ai/glm-4.6)
Price: $1.07/1M | Context: 202K
Underrated. Good balance of price and capability. Handles long context well.
Good for: Documentation, explanations, medium complexity tasks Bad for: Cutting-edge reasoning
Qwen3 VL (qwen/qwen3-vl-235b-a22b-instruct)
Price: $1.06/1M | Context: 131K
Vision + code combo. Best for when you need to work with screenshots, designs, or diagrams.
Good for: UI work from screenshots, diagram understanding, visual debugging Bad for: Extended reasoning (no reasoning capability)
Pricing Reality Check
Let's do real math.
Average coding session: ~50K tokens (input + output)
| Model | Cost per 50K tokens |
|---|---|
| MiniMax M2 | $0.03 |
| Grok Code Fast | $0.04 |
| GLM 4.6 | $0.05 |
| Qwen3 VL | $0.05 |
| GPT-5.1 Codex | $0.28 |
| Gemini 3 Pro | $0.35 |
For most tasks, we're talking cents. Don't obsess over pricing unless you're doing high-volume automation.
Model Selection Strategy
For experiments: Start cheap (MiniMax M2). See if it works.
For important code: Use a capable model (Grok, Codex). It's still cheap.
For architecture decisions: Go premium (Gemini 3 Pro). Context and reasoning matter.
For automation: Pick the cheapest that works reliably for your task.
Custom Models
See a model on OpenRouter that's not in our list? Use it anyway:
claudish --model anthropic/claude-sonnet-4.5 "your prompt"
claudish --model mistralai/mistral-large-2411 "your prompt"
Any valid OpenRouter model ID works.
Force Update Model List
The model cache updates automatically every 2 days. Force it:
claudish --top-models --force-update
Next
- Model Mapping - Use different models for different Claude Code roles
- Cost Tracking - Monitor your spending