Set a budget and how you work. Everything re-ranks as you type, and the link in your address bar saves the whole setup.
These numbers may be out of date — the daily check has not run recently.
The share link reproduces your exact setup.
Every stack that fits your settings. The gold line is the best you can buy at each price. Past the ceiling you can't use the extra hours, so the ranking weighs model quality and provider mix instead.
Hover any dot to see the stack. Click to pin it here.
Your #1 stack, split by plan, at your off-peak share. Striped = extra hours Z.ai's half-price off-peak window gives you.
The #1 stack at each budget, with your other settings unchanged. Click a row to switch to it.
Weekly building hours on a typical week (the 50th percentile). Faded icons have a cheaper or better alternative. Click a row for the sources and haircut notes.
Catalog updates from the daily check, with the source each one came from.
Plain rules. Every number traces to catalog.js.
Vendor banners are ceilings. We store what people actually get: a bad week (p10), a typical week (p50) and a good week (p90), cut down from the banners toward lived reports.
Each plan is rated cheap, strong or frontier for planning, building and admin. There are no leaderboard scores here. Stacks that miss the minimum you set are dropped.
Two plans from one vendor share one reset clock, so one bad morning hits both. Stacks spread across providers score higher, and a second model family is there when one gets stuck.
Z.ai is the only vendor that publishes a night rate (credits cost half outside Mon–Fri 07:00–11:00 London). Nobody else's hours move with the off-peak slider.
Useful hours are capped at 1.6× your keyboard time. Past that, a bigger plan buys nothing, so the ranking weighs quality and provider mix instead.
Some links are referral links. The scorer never sees them. Paused plans stay listed with a tag so a stale blog post can't sell you a dead plan.
Put your budget into the stack builder. At $200–300 the top stacks usually pair a frontier-grade planning plan with a high-hours building plan, plus one or two more providers so you have a second reset clock and a second model family to try. Every pick is recomputed from the catalog as you change settings.
One vendor = one clock and one model family: a bad morning or a thin week hits every seat at once. Mixing pools spreads reset clocks and quality grades, and no single plan is Pareto-best at every budget. Finding the best mix is the optimizer's job.
Z.ai publishes a 0.5× credit rate outside Mon–Fri 07:00–11:00 London and all weekend. It stretches hours; it does not raise concurrency or enlarge the 5h credit budget. It's the only published night rate in the catalog. Everyone else charges the same all day.
Vendor banners are ceilings. Each plan's hours are cut toward what people actually see in lived reports. For example, Claude implement hours are a haircut of official Sonnet numbers, and Fable plan hours are never added on top because they share one pool.
Every plan carries a source tag: official / official-clock / community / single-report / guessed. Weakly-sourced plans are marked amber in the table and flagged as the weakest link whenever they appear in a recommended stack.
Every plotted number traces to catalog.js; every rule to the README's honesty section. There's no SWE-bench and no leaderboard axis. Quality is a 1–3 role grade, never a benchmark score.
No vendor paid to be here. Some plan links are referral links: the site may earn credits or a commission if you subscribe, and you sometimes get a discount. The scorer never sees them. Paused plans stay listed, tagged, and never appear as buyable picks.
So you don't buy a dead plan from a stale blog post. Paused plans are tagged red and sort below every buyable stack; they're hidden unless you ask for them.
Free · open source · Windows
Once you run three subscriptions, the hard part is knowing which one still has room. AI Usage Tracker puts every plan's 5-hour and weekly quota in one tray window.