ai-model-selector（AI 模型选择器）

AI Model Selector Skill

Comprehensive guide to selecting and implementing AI models. Updated January 2026.

Quick Decision Tree

What's your primary need?
│
├─► CODING/AGENTIC TASKS
│   ├─► Best quality → Claude Sonnet 4.5 or GPT-5.2-Codex
│   ├─► Complex reasoning → Claude Opus 4.5 (with effort param)
│   └─► Budget → DeepSeek-chat ($0.28/1M input)
│
├─► REASONING/MATH/SCIENCE
│   ├─► Maximum intelligence → GPT-5.2 Pro or Claude Opus 4.5
│   ├─► Good balance → GPT-5.2 (xhigh effort) or Gemini 2.5 Pro
│   └─► Budget → DeepSeek-reasoner (visible CoT)
│
├─► LONG DOCUMENTS (>200K tokens)
│   ├─► Up to 1M tokens → Claude Sonnet 4.5 (beta) or Gemini 2.5 Pro
│   ├─► Up to 400K → GPT-5.2
│   └─► Budget → DeepSeek-chat (128K)
│
├─► HIGH-VOLUME/LOW-LATENCY
│   ├─► Best speed → Claude Haiku 4.5
│   ├─► Cheapest → Gemini 2.5 Flash-Lite ($0.10/$0.40)
│   └─► Free tier → Gemini via AI Studio
│
├─► EMBEDDINGS/RAG
│   ├─► Best quality → Voyage 3.5 or voyage-3-large
│   ├─► Code-specific → voyage-code-3
│   ├─► Budget → text-embedding-3-small ($0.02/1M)
│   └─► Free → gemini-embedding-001
│
└─► MULTIMODAL (images/audio/video)
    ├─► Images → GPT-4o, Gemini 2.5 Pro/Flash, Claude 4.5
    ├─► Image generation → GPT Image 1, Imagen 4.0
    └─► Video generation → Veo 3.1

Model Quick Reference (January 2026)

Flagship Models

Budget Models

Critical Gotchas

⚠️ GPT-5.x / O-series Don't Support These Parameters:

// WRONG - will error on GPT-5.2, o3, o4-mini
{
  temperature: 0.7,      // ❌ Not supported
  top_p: 0.9,            // ❌ Not supported
  max_tokens: 4096,      // ❌ Use max_completion_tokens
}

// CORRECT
{
  reasoning: { effort: "high" },  // none, low, medium, high, xhigh
  text: { verbosity: "medium" },  // low, medium, high
  max_completion_tokens: 4096
}

⚠️ Claude Opus 4.1 vs 4.5 Pricing

• Opus 4.1: $15/$75 per 1M tokens (legacy pricing)
• Opus 4.5: $5/$25 per 1M tokens (66% cheaper, better quality!)
• Always use Opus 4.5 for new projects

⚠️ Long Context Premium Pricing (Claude Sonnet)

• ≤200K tokens: $3/$15 per 1M
• 200K tokens: $6/$22.50 per 1M (automatic)

Detailed Documentation

• OpenAI Models - GPT-5.2, GPT-5.1, GPT-4.1, o-series
• Anthropic Models - Claude Opus/Sonnet/Haiku 4.5
• Google Models - Gemini 2.5/3 Pro/Flash
• DeepSeek Models - V3.2, R1 (cheapest option)
• Embedding Models - Voyage, OpenAI, Gemini

Use Case Guides

• Coding Tasks - Agentic coding, code review
• Reasoning Tasks - Math, science, planning
• RAG Pipelines - Embeddings, retrieval, generation

Cost Optimization

Batch API (50% off)

All major providers offer batch processing for non-urgent tasks:

• OpenAI: 50% off all models
• Anthropic: 50% off all models
• DeepSeek: 33% off

Prompt Caching

• Claude: 90% savings on cache reads
• OpenAI: 90% savings on cached inputs
• DeepSeek: Automatic caching, 90% off hits

Model Cascading

Route simple queries to cheap models, complex to expensive:

Simple question → Haiku 4.5 ($1/$5)
Complex task → Sonnet 4.5 ($3/$15)
Hardest problems → Opus 4.5 ($5/$25)

API Code Templates

OpenAI (GPT-5.2)

const response = await fetch("https://api.openai.com/v1/responses", {
  method: "POST",
  headers: {
    "Authorization": `Bearer ${OPENAI_API_KEY}`,
    "Content-Type": "application/json"
  },
  body: JSON.stringify({
    model: "gpt-5.2",
    input: [{ role: "user", content: "Hello" }],
    reasoning: { effort: "medium" }
  })
});

Anthropic (Claude)

import Anthropic from '@anthropic-ai/sdk';
const anthropic = new Anthropic();

const response = await anthropic.messages.create({
  model: "claude-sonnet-4-5-20250929",
  max_tokens: 4096,
  messages: [{ role: "user", content: "Hello" }]
});

Google (Gemini)

import { GoogleGenerativeAI } from "@google/generative-ai";
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
const model = genAI.getGenerativeModel({ model: "gemini-2.5-flash" });

const result = await model.generateContent("Hello");

DeepSeek (OpenAI-compatible)

import OpenAI from 'openai';
const client = new OpenAI({
  baseURL: 'https://api.deepseek.com',
  apiKey: process.env.DEEPSEEK_API_KEY
});

const response = await client.chat.completions.create({
  model: 'deepseek-chat',
  messages: [{ role: 'user', content: 'Hello' }]
});

Benchmark Reference (January 2026)

SWE-bench Verified (Coding)

1. Claude Opus 4.5: 80.9%
2. GPT-5.1-Codex-Max: 77.9%
3. Claude Sonnet 4.5: 77.2%
4. GPT-5.2-Codex: ~78% (est.)

AIME 2025 (Math)

1. GPT-5.2 (xhigh): 100%
2. o3: 90%+
3. Claude Opus 4.5: High 80s%
4. DeepSeek R1: 79.8%

GPQA Diamond (Science)

1. GPT-5.2: ~92-93%
2. Claude Opus 4.5: ~85%+

*Last updated: January 28, 2026* *Sources: Official documentation from OpenAI, Anthropic, Google, DeepSeek*