LLM Cost Optimizer

Analyze LLM API usage patterns and suggest optimizations to reduce costs by 50%+.

Quick Start

Run the analysis script against a usage log JSON file:

python3 scripts/analyze.py --input usage_logs/

Input Format

Usage logs are JSON arrays of objects with:

• timestamp (ISO 8601)
• model (e.g. "gpt-4o", "claude-sonnet-4-20250514")
• input_tokens (int)
• output_tokens (int)
• task_type (optional: "general", "coding", "classification", "extraction", "summary", "creative", "analysis")
• session_id (optional)
• cached (optional bool)

Optimization Strategies

1. Downgrade: Find simple calls (low output, basic task types) on premium models and suggest cheaper alternatives
2. Cache: Detect repeated prompt patterns and recommend prompt caching
3. Compress: Find verbose outputs and suggest prompt tuning for conciseness

Supported Models & Pricing

All pricing is per 1M tokens (input/output). Rates are in the PRICING dict in scripts/analyze.py.

Model Tiers

• Premium: gpt-4o, claude-sonnet-4, gemini-2.5-pro
• Mid: gpt-4.1-mini, claude-haiku-3-5, gemini-2.5-flash, deepseek-v3
• Budget: gpt-4.1-nano, deepseek-r1, llama-4-maverick

Downgrade suggestions map premium → mid → budget for simple tasks.