Cost Optimization

Strategy 1: Prompt Caching (50-90% Savings)

The single biggest cost lever. If you send the same system prompt across many requests, prompt caching avoids re-processing those tokens.

Provider	Savings	Trigger
OpenAI / Azure	50%	Automatic, prefix >= 1024 tokens
Anthropic	90%	Gateway auto-injects `cache_control` (system prompt >= 3000 chars)
DeepSeek	90%	Automatic disk caching
Gemini 2.5+	90%	Implicit caching

Impact example: A Claude Code user sending 50 requests/day with 20K-token system prompts saves ~$81/month on input tokens alone.

See the Prompt Caching guide for details.

Strategy 2: Optimal Routing

The same model is often available through multiple providers. Chuizi.AI automatically routes to the optimal provider for best cost and performance. You benefit from this without any configuration -- just use the model name (e.g., deepseek/deepseek-chat) and the gateway handles provider selection.

For models available through multiple providers, this automatic routing can result in significant cost savings compared to using a single provider directly.

Strategy 3: Model Tiering

Not every task needs the most capable model. Match model capability to task difficulty:

Task Type	Recommended Model	Input Price	Output Price
Classification, tagging, simple extraction	`openai/gpt-4.1-nano`	$0.10/M	$0.40/M
Summarization, translation, Q&A	`openai/gpt-4.1-mini`	$0.40/M	$1.60/M
Code generation, analysis	`anthropic/claude-sonnet-4-6`	$3.00/M	$15.00/M
Complex reasoning, research	`openai/gpt-5`	$2.63/M	$15.00/M
Hardest problems, agentic workflows	`anthropic/claude-opus-4-6`	$15.00/M	$75.00/M

Cost difference: Processing 1M input + 100K output tokens:

With GPT-4.1-nano: $0.14
With Claude Opus 4-6: $22.50
Ratio: 160x

Use a routing strategy in your application: classify the task first with a cheap model, then dispatch to the appropriate tier.

Strategy 4: Reduce Token Usage

Set `max_tokens` explicitly

Without max_tokens, some models generate until their context limit. Always set a reasonable cap:

config.json

json

{
  "model": "openai/gpt-4.1",
  "messages": [{"role": "user", "content": "Translate to French: Hello"}],
  "max_tokens": 100
}

Trim conversation history

Long conversation histories inflate input tokens. Keep only the messages the model needs:

example.py

python

# Instead of sending the entire history:
messages = conversation_history[-10:]  # Keep last 10 turns

# Or summarize older messages:
messages = [
    {"role": "system", "content": "Previous context summary: ..."},
    *recent_messages,
]

Use `detail: "low"` for vision

When analyzing images, use low detail mode unless you need OCR or fine-grained analysis: