Comprehensive comparison of AI models covered in the Prompt Library — pricing, capabilities, latency, rate limits, and selection guidance.

⚠️ Pricing and model availability change frequently. This guide reflects information as of April 2026. Always verify current rates on each provider's pricing page before making cost projections.

• Claude (Anthropic)
• GPT (OpenAI)
• Gemini (Google)
• Llama (Meta)
• Mistral
• Pricing Comparison Table
• Benchmark Comparisons by Task
• Latency Comparison
• Rate Limits
• Decision Tree: Which Model to Use
• Cost Optimization Strategies
• Retired Models

Claude models excel at instruction following, long-context analysis, creative writing, and coding. They respond exceptionally well to structured prompts using XML tags.

XML Tags — Claude treats XML tags as semantic containers. High-value patterns:

<instructions>Primary task definition</instructions>
<context>Background information</context>
<constraints>Hard rules that override other instructions</constraints>
<examples>Input/output examples</examples>

Extended Thinking — Available on Opus 4.6, allows the model to "think" through complex problems before responding. Triggers deeper reasoning at the cost of higher latency and tokens.

Prefill Technique — Start Claude's response to steer format and tone:

{
  "messages": [
    {"role": "user", "content": "Analyze this data..."},
    {"role": "assistant", "content": "{\"analysis\":"}
  ]
}

• Opus 4.6 — When you need the absolute best reasoning, processing documents over 200K tokens, or running multi-agent workflows
• Sonnet 4.6 — The workhorse — use for 90% of tasks (coding, analysis, writing, tool use)
• Haiku 4.5 — High-volume classification, routing, quick extractions, or where latency/cost matters most

GPT models provide strong general-purpose performance, structured outputs, and function calling.

JSON Mode — Force output to be valid JSON:

{
  "response_format": { "type": "json_object" }
}

Structured Outputs — Define an exact JSON schema the model must follow:

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "analysis",
      "schema": {
        "type": "object",
        "properties": {
          "summary": { "type": "string" },
          "score": { "type": "number" }
        },
        "required": ["summary", "score"]
      }
    }
  }
}

Function Calling — Define tools the model can invoke, enabling agent-like behavior with structured tool use.

Gemini models offer native multimodal capabilities, Google Search grounding, and code execution.

Multimodal Input — Native support for images, video, and audio in prompts. No need to describe visual content in text.

Google Search Grounding — Ground responses in real-time web search results for up-to-date information.

Code Execution — The model can execute code as part of its response, useful for data analysis and computation tasks.

Open-weight models you can self-host. No API costs — you pay only for compute.

• Open-weight — download and run on your own infrastructure
• No API costs — only compute costs (GPU rental or owned hardware)
• Fine-tuning — customize for your specific domain
• MoE architecture (Llama 4) — high parameter count with efficient inference
• Privacy — data never leaves your infrastructure

• Data must stay on your infrastructure (regulatory, privacy)
• High-volume inference where API costs would be prohibitive
• Fine-tuning for specialized domains
• Edge/mobile deployment (Llama 3.2)
• Research and experimentation

European AI lab offering a range of models from tiny to frontier, with excellent price-to-performance ratios.

• Best price-to-performance: Mistral Medium 3 at $0.40/M input offers GPT-4-class quality at 1/5 the cost
• Multilingual excellence: Strong across European and global languages
• MoE architecture: Efficient inference with high parameter counts
• Code specialization: Codestral and Devstral for development workflows
• EU data sovereignty: European hosting options

Prices per million tokens (April 2026):

*Llama models are open-weight — you pay only for compute (GPU hosting/rental).

Batch API Discounts:

• OpenAI Batch API: 50% off (24hr turnaround)
• Anthropic Batches API: 50% off (24hr turnaround)
• Google Batch API: varies by model

Practitioner's view of relative performance per task category:

Approximate ranges for typical requests (varies by region, load, and prompt length):

TTFT = Time To First Token. These are rough medians for guidance, not SLAs.

RPM = requests per minute, RPD = requests per day. Limits vary by model within each provider.

START: What is your primary requirement?

[Data must stay on your infrastructure?]
  YES → Llama 3.3 70B (quality) or Llama 3.2 3B (edge/mobile)
  NO → continue

[Processing video or audio natively?]
  YES → Gemini 3 Pro (native video/audio)
  NO → continue

[Documents exceeding 200K tokens?]
  YES → Claude Opus 4.6 (1M) or Llama 4 Scout (10M)
  NO → continue

[Need guaranteed JSON schema compliance?]
  YES → GPT-5.4 with structured outputs
  NO → continue

[Complex reasoning or long-form writing?]
  YES → Claude Sonnet (value) or Opus (maximum quality)
  NO → continue

[High-volume, cost-sensitive (>10K req/day)?]
  YES → How complex?
    Simple → Gemini Flash or Mistral Medium 3
    Moderate → Claude Haiku 4.5
    Complex → Claude Sonnet with batching
  NO → continue

[Code generation or review?]
  YES → Claude Sonnet 4.6, GPT-5.3-Codex, or Codestral
  NO → continue

[Default / general purpose]
  Budget → Mistral Medium 3
  Quality → Claude Sonnet 4.6 or GPT-5.4
  Maximum → Claude Opus 4.6

Use a cheap model to classify request complexity, then route to the appropriate model:

# Classify with a cheap model
classification = cheap_model.classify(
    request, categories=["simple", "moderate", "complex"]
)

model_map = {
    "simple":   "gemini-flash",      # lowest cost
    "moderate": "claude-haiku",      # balanced
    "complex":  "claude-sonnet",     # highest quality
}
model = model_map[classification]

response = cheap_model.generate(prompt)

if not passes_quality_check(response):
    response = expensive_model.generate(prompt)  # escalate

task_routing = {
    "classification": "gemini-flash",
    "extraction":     "mistral-medium",
    "summarization":  "claude-haiku",
    "reasoning":      "claude-sonnet",
    "code":           "codestral",
}

A well-implemented router saves 60–80% compared to sending everything to the most expensive model.

Do not use these in new projects:

Navigation: ← Prompting Techniques  |  Tools: Linter, Optimizer, Recommender →