Which subscription buys the most usage per dollar?
MTok means 1 million tokens, how providers measure usage. Each plan is ranked by the million tokens you get per dollar, under the coding scenario and reference model you pick.
Podium (TL;DR)
Coding Scenario
Pick the day of work your subscription must cover. Every rank below reflects it.
· active / total
How to Get Started
What you are buying
Coding tools like OpenCode, Codex, Claude Code, and Pi send your prompts to an AI provider. The provider bills for the work. The providers in this table replace the per-token bill with a flat monthly plan.
- API
- The paid endpoint a coding tool sends prompts to. OpenAI and Anthropic run the best-known ones; the providers in this table serve compatible endpoints.
- Tokens
- The words and symbols a model reads and writes; one token is about 3/4 of a word. Providers measure usage in millions of tokens (MTok).
- Credit pool
- The monthly dollar balance a credit plan holds. Each request spends it at the model's own rate.
- Cache reads
- Prompts the provider serves from its cache instead of fresh compute. They bill far below fresh input.
Connect a plan in five minutes
- Pick a provider in the table below; check its badge and notes.
- Sign up on the provider's page and copy your API key.
- Paste the key into your coding tool. Each tool documents the step:
How I test
I run real coding agent workloads against every provider in this table. The badge sums up what happened: approved kept its promise, ok'd works with a stated flaw, shaky served unstable models, failed broke its core promise, and wow'd beat every tested alternative on a measured trait. The numbered notes under each plan carry the details.
Credit Plan Providers
| Models | Notes | |||||||
|---|---|---|---|---|---|---|---|---|
|
|
· | - MTok/$1est. source |
Unlimited Plans
Unlimited plans cannot be priced per token; this table shows what each provider actually limits. The rank is an editorial pick: usable models first, then published capacity, zero-data retention, and price. Plans that do not serve the selected reference model drop to the bottom, unranked. Providers that serve none of the reference models always sit there. Every plan here works with any harness.
First-Party Labs
The rank is editorial: plans with the most usable tokens per dollar.
Frequently Asked Questions
What does $1 → $X mean?
It is the credit multiplier: how many dollars of usage one dollar of the plan buys. A $20 plan with $1 → $6 holds about $120 of usage at the published rates.
What is the difference between the two rankings?
The catalog view — the model set to None — ranks every plan by credit value per dollar, whatever model you will run. The scenario view ranks plans by the million tokens they serve under your workload for the selected reference model. Use the catalog view to compare plans; use the scenario view to compare under your daily workload.
Do these plans work with my coding tool?
Check the Harness column. Any means the plan works with OpenCode, Codex, Claude Code, Pi, and other agent tools. Proprietary means the provider's own tool only.
Is my code kept private?
Check the ZDR column. Yes means zero-data retention: the provider does not store your prompts or code. Unverified means no policy states it, and No means the provider may keep data.
What happens when I hit the limit?
Every plan's notes state the windows and caps. Credit plans stop or bill overage at a discount; cap plans throttle or queue. The Rate limit logic column on the Unlimited table carries the exact numbers.
Why is a plan dimmed at the bottom of the table?
Dimmed plans do not qualify for a rank: the provider failed testing, serves none of the reference models, or does not serve the model you picked.
What does the tested badge mean?
It sums up hands-on testing. The How to Get Started section explains each state, and the numbered notes under each plan carry the details.
Where does the data come from?
All data comes from live web research: the providers' own pricing and policy pages, plus hands-on testing.