Coding subscriptions are priced in deliberately incompatible units: Anthropic counts messages over a five-hour window, Cursor sells opaque credits, the Chinese labs sell prompts or requests, and some plans publish a hard weekly cap on top. You cannot compare them directly — which is exactly why “best value” listicles hand-wave. We don’t.
Step 1 — normalize the native unit to tokens
Each plan’s allowance is expressed in its real native unit (prompts, requests, messages, credits, or raw tokens) over its real refresh window. We convert that to an estimated token allowance using a single, adjustable assumption — tokens per interaction — and per-plan scaling where a “request” is materially smaller than a “prompt.” Token-native plans skip this step entirely.
Step 2 — extrapolate to a monthly basis, then clip to caps
We scale the per-window allowance to a week using your active hours per week, then to a month. Where a vendor publishes a hard weekly cap, we apply it as a ceiling — so a plan can’t look better than its own published limit allows.
Step 3 — the headline numbers
- Effective $/M token — monthly price ÷ included monthly tokens. The cost lens.
- API-equivalent multiple — what that same usage would cost at direct API pricing, divided by the subscription price. Above 1× means the plan beats the API; below 1× means you’d be better off paying per token.
- Value — tokens per dollar weighted by the model’s Artificial Analysis index, so cheap-but-weak and strong-but-pricey both get their due.
Where the estimates stop
These are estimates from your assumptions, not vendor guarantees. Real usage varies with prompt size, caching, and how a vendor meters under load. We show the assumptions on-screen and make them adjustable precisely so you can stress-test the answer rather than take it on faith. Every price and index links to its source; the snapshot is dated.