If you’ve ever tried to compare AI coding plans, you know the pain. One provider advertises “unlimited” tokens with vague fair-use limits. Another sells credits that convert to tokens at rates that depend on cache hit ratios you can’t predict. A third changes its pricing structure every few weeks. By the time you’ve built a spreadsheet, half the numbers are stale.
Real API Pricing is a GitHub project that cuts through this mess with a single, brutally simple metric: monthly subscription fee ÷ monthly usable tokens. It’s maintained by FeiZhuLulu, and it’s already attracted attention on GitHub Trending and X for its clear-eyed approach to a problem that every developer paying for AI tools faces.
The Core Problem: Incomparable Pricing Units
AI providers deliberately make comparison hard. Some charge per token, some per credit, some per message, some per “pool” that refreshes weekly or monthly. Kimi’s monthly pool is five times its weekly pool. GLM Coding Plan uses weekly credits with separate cache, input, and output coefficients. StepFun prices in CNY credits (1M Credit = ¥1) while its international site uses USD sticker prices that don’t match.
Real API Pricing solves this by converting everything to a common unit: dollars per million tokens ($/MTok). The project’s central formula is stated plainly in the README: Real unit price = monthly subscription fee ÷ monthly usable tokens.
That sounds simple, but executing it rigorously requires a lot of decisions. The project documents every one of them.
A Standard Workload Makes Comparison Possible
The trickiest part of normalizing AI pricing is that token types aren’t interchangeable. Cache reads are cheap (sometimes free), fresh input costs more, and output costs the most. A plan that looks generous on paper might be stingy if you’re mostly generating output rather than reading cached context.
To handle this, Real API Pricing adopts a project-wide standard workload: 97.5% cache reads, 2.15% fresh input, and 0.35% output. This is explicitly framed as a comparison convention, not a claim about any provider’s actual workload. It’s the pricing equivalent of a benchmark—a fixed reference point that lets you rank options apples-to-apples.
When the project already has total-token measurements from dashboards, local logs, saturation tests, or official absolute-token tables, it doesn’t re-normalize them. It only applies the standard workload when converting dollar/credit pools or three-part token prices.
What’s Actually in the Dataset
As of the September 9, 2026 snapshot, the project covers:
- 202 adopted plan × model points
- 188 subscription points with monthly allowance
- 13 metered API baselines
- Scored points across Code Arena (136), Agent Arena (140), AA Intelligence (173), AA Coding Agent (71), OpenDesign Arena (70), and Terminal-Bench 4.0 (70)
Each row represents one plan × actual served model. Crucially, allowances for different models under the same plan are treated as alternatives, not additive—you can’t sum Claude and GPT allowances on a plan that lets you switch between them.
The data is available as CSV and JSON, and the project maintains a detailed research archive with dated evidence files.
Pareto Charts: Seeing the Frontier
The most visually striking output is a set of Pareto charts, one per leaderboard. These plot each model’s benchmark score on the Y-axis against its real API price on the X-axis, then draw the Pareto frontier—the set of options where no other option is both cheaper and better.
Subscriptions and metered APIs compete on the same frontier. The charts use a logarithmic price axis, so cheaper points sit farther right. The project includes interactive HTML versions with a reasoning-effort selector, allowing you to explore how different configurations shift the frontier.
One notable edge case: SWE-2 · Devin Pro is unmetered during a promotional period, so its real price is shown as ≈$0/MTok on a dedicated axis slot, making it the cheapest frontier point. The project flags this explicitly as a promotional price that must be re-evaluated when the promotion ends.
Who This Is For
Real API Pricing is useful for three audiences:
- Individual developers choosing between $20, $100, or $200/month plans who want to know which gives the most usable tokens per dollar.
- Teams evaluating enterprise tiers where “unlimited” claims hide significant variation in actual throughput.
- Analysts and researchers studying the economics of AI infrastructure, who need transparent, reproducible pricing data with documented methodology.
The project’s data sources include Awesome Coding Plan (CC BY 4.0) and the Caijing article 《Token经济，中国账本》, with full attribution in SOURCES.md.
Access the Data
- Interactive website: real-api-pricing.vercel.app
- Repository: github.com/FeiZhuLulu/real-api-pricing
- Charts: Available in English and Chinese, SVG and PNG formats
- Raw data: adopted.csv and derived/points.csv
The Bottom Line
Real API Pricing doesn’t pretend to tell you which model is “best”—that depends on your workload, your tolerance for latency, and the quality of outputs for your specific tasks. What it does is give you a defensible, transparent answer to a narrower but still crucial question: How many usable tokens do I actually get for my monthly fee, and how does that compare across every major option?
In a market where pricing pages are designed to obscure rather than clarify, that’s a genuinely useful contribution. The project is open-source under MIT, the methodology is documented, and the data is downloadable. If you’re paying for AI tools, it’s worth a look.