For teams
Buying seats for Claude Code, Codex, or Grok Build? Here is steady, dated data you can check yourself before you sign.
The three questions buyers ask
Which is fastest at the effort we would use?
The live board ranks the four featured models — Fable 5, Opus 5, GPT-5.6 Sol, and Grok 4.6 — on wall tok/s, the one speed unit that is fair across all three apps. Every model runs at the top of its own effort scale.
How often does it rate-limit?
We hit the same paid plans every hour, all day. A limit shows up on our clock before it shows up in your sprint. Today counts every limit from the last 24 hours, and Status shows the recent ones.
Does it finish the job every round?
Every round is graded: a run only counts when the answer is right. We publish clean rounds out of rounds tried, so a tool that is quick but flaky cannot hide behind a good median.
A buyer’s check, in order
Twenty minutes across these five pages will tell you more than a vendor deck.
- Compare the two you are choosing between
Same 24-hour window, same unit, side by side. Rank on wall tok/s, and glance at clean runs — a number from 3 rounds can swing, one from 90 barely moves.
- Open the page for the model you would actually use
One page per pin: write speed, time to first word, clean rounds, rate limits, and how it moved since yesterday. A team seat runs one model, not an average of four.
- Check this week, not just the average
Logins, rate limits, and the exact app versions behind today’s numbers. We measure on a fixed clock, so a vendor change shows up the day it ships.
- Open the samples behind the summary
Every run we saved, every fail with a label, and the tokens our lab spent. If a board looks too good, this is where you go to doubt it.
- Read the fine print before you quote it
Exact prompts, exact flags, pinned models, and the test kit version — boards only show the current kit, v6. Do not put a number in a deck before you read this page.
How to read this
Every figure on this site is one we measured ourselves on our own machine. None of it is an official vendor benchmark. Read How we test before you quote a number in a buying decision.
What we never do
- We never sell a spot on a board. No vendor has ever paid us a cent, for placement or anything else.
- We never gate the data. No form, no demo call, no email wall between you and any number on this site.
- We never blend two different measurements into one winner score that could be quoted out of context.
The longer story is on About.
Honest limits
Read these before you put a figure in a business case.
- One machine, one network
- Every number comes from one dedicated lab machine on a normal home connection, using the app versions listed on Status. Your office setup will not match ours exactly.
- Short jobs only
- We time a tiny ping and a 1-to-120 count. We do not yet measure long agent work in a real repository.
- No pricing model
- We publish the tokens our lab spent, not what a seat will cost your team. Plan pricing is the vendor’s page, not ours.
- Not a contract number
- These are our measurements, not official vendor numbers, and no vendor promises them. Use them to compare tools and spot a bad week. Do not paste one into a contract or an SLA.
Talk to us
Use the contact link in the footer if you need a model we do not track yet, or want a number checked before a big buy. A person answers, and we will tell you plainly if the data cannot support the question. The raw numbers are on the Developers page as JSON feeds — no key, no account.