FeiZhuLulu/real-api-pricing

What it solves

It provides a standardized way to compare the actual cost of AI model subscriptions and metered APIs. Because different providers use varying pricing models (monthly fees, credit pools, and different rates for input, output, and cache tokens), it is difficult for users to determine which plan offers the best value for their specific usage.

How it works

The project calculates a "Real unit price" by dividing the monthly subscription fee by the monthly usable tokens. To make comparisons fair across different pricing structures, it applies a project-wide standard workload (97.5% cache reads, 2.15% fresh input, and 0.35% output) to convert various credit pools and three-part token prices into a single comparable metric. It then maps these costs against performance scores from various leaderboards (such as Code Arena, Agent Arena, and AA Intelligence) to create Pareto charts that visualize the best trade-offs between cost and intelligence.

Who it’s for

Developers and AI power users who want to optimize their spending on LLM subscriptions or API usage based on actual token throughput and model performance.

Highlights

• Standardized Workload: Uses a consistent token mix to normalize disparate pricing models into a single $/MTok scale.
• Pareto Frontier Analysis: Visualizes the cost-performance trade-off using data from multiple AI leaderboards.
• Comprehensive Coverage: Tracks hundreds of subscription plan and model combinations, including both metered APIs and monthly subscriptions.
• Multi-dimensional Data: Provides data snapshots in CSV and JSON formats, including detailed benchmark configurations and plan-to-configuration mappings.