Skip to content
BetaBenchAlert v0.1 is in Beta.Numbers are real, but pages and rules can still change.See what changed

Lab health

Is the lab healthy?

The machine behind every number on this site: is it awake, is it still signed in, and what did the last round of checks do?

Lab is runningSigned in and ready

Running means the last round started under 180 minutes ago. A few rate limits a day is normal on paid plans.

If the lab is behind, read the boards as stale and check back. Reading this site never starts a test — the schedule is ours alone.

Next check

A new round starts every hour, day and night. The next one is at 4:00 AM.

How fresh is this?

When the lab last ran, and how far behind it is right now.

Last speed round

4m ago · 13 saved · 0 skipped · 2 failed · took 4m 46s

The hourly ping and speed sweep

Most recent problem

overloaded

The newest error label in the last 24 hours

What counts on the boards

Test kit v0.2 only · the answer must pass our grader · leftover plugins, skills, or hooks are named on the sample

The rule the live boards apply to every sample

Where each plan's meter stands

How full our own usage window is on each paid plan right now, when it resets, and how solid today's reading is. This is our account, not yours.

See what a full window is worth
Claude Max 20x
71% of the window usedGrade · Gaps

Resets Thu 04:00 UTC · 16 readings that day · read 3h ago

ChatGPT Pro ($200, 20×)
4.0% of the window usedGrade · Gaps

Resets Mon 00:41 UTC · 1687 readings that day · read 3h ago

SuperGrok Heavy
37% of the window usedGrade · Gaps

Resets Sun 17:53 UTC · 12 readings that day · read 3h ago

We read our own account's meter from each vendor's usage endpoint, a few times an hour. It costs no model calls, and it says nothing about your account.

24-hour tape

The last 24 hours

Every ping, lined up round by round. This is the app waking up — not a writing-speed test.

Claude Code

Anthropic

Fastest this round

3.21s

Start-to-finish this round

100%

Pings that worked in 24h

31% fasterthan its usual 24-hour time

Start-to-finish · 24 hours

Usual
4.69s
Slowest
8.36s
Finished
24/24

Codex

OpenAI

Live

5.94s

Start-to-finish this round

100%

Pings that worked in 24h

7% fasterthan its usual 24-hour time

Start-to-finish · 24 hours

Usual
6.40s
Slowest
18.7s
Finished
21/21

Grok Build

xAI

Live

4.63s

Start-to-finish this round

100%

Pings that worked in 24h

12% fasterthan its usual 24-hour time

Start-to-finish · 24 hours

Usual
5.28s
Slowest
17.7s
Finished
24/24

How to read this

Ping is a tiny ok check. Wall time is start to finish — lower is better. In the chart, a red hashed stub failed, and a bar that fades at the top ran past the scale — hover it for the real number. Writing speed lives on the home leaderboard, in visible answer tok/s.

Rounds with a check
24 of 24
Each round is hour
Checks saved
69
0 failed
Checks that worked
100.0%
Finished with no error

Wall time · last 24 hours

One bar per agent, every hour

Every column is one slot on the same clock, so the bars line up across agents.

Claude Codemedian 4.69sCodexmedian 6.40sGrok Buildmedian 5.28sShorter bars are faster
wall05s10s15s20s
4amAug 236am8am10am12pm2pm4pm6pm8pm10pm12amAug 242am3am

How to read this

Each one-hour slot has one bar per agent, side by side. Height is how long the ping took, so a shorter bar is faster. A red hatched stub is a probe that failed. A bar that fades at the top ran past the scale — hover it for the real number. A thin flat line means no probe ran that slot.

Full record — every slot, every number
HourClaude CodeCodexGrok Build
Wall timeTok/sWall timeTok/sWall timeTok/s
Aug 24 3:00am3.21s12.455.94s0.844.63s3.67
Aug 24 2:00am3.45s22.345.78s0.875.21s3.26
Aug 24 1:00am4.75s13.8918.7s0.276.90s4.35
Aug 24 12:00am4.69s25.187.06s0.716.15s4.88
Aug 23 11:00pm4.14s34.276.20s0.814.46s3.81
Aug 23 10:00pm5.63s6.9318.5s0.276.39s4.70
Aug 23 9:00pm4.58s25.0912.0s0.424.22s7.11
Aug 23 8:00pm4.83s15.9311.3s0.4417.7s0.96
Aug 23 7:00pm4.50s18.899.60s0.524.70s6.18
Aug 23 6:00pm3.19s20.70no sample8.01s3.62
Aug 23 5:00pm3.48s21.57no sample6.61s4.39
Aug 23 4:00pm4.37s17.84no sample7.28s3.98
Aug 23 3:00pm4.94s14.778.52s0.595.80s2.93
Aug 23 2:00pm5.14s20.646.40s0.785.19s5.59
Aug 23 1:00pm4.69s22.176.54s0.765.23s3.25
Aug 23 12:00pm4.73s13.1210.1s0.504.79s3.76
Aug 23 11:00am4.90s19.614.73s1.0615.3s1.90
Aug 23 10:00am7.62s16.414.72s1.065.26s5.52
Aug 23 9:00am8.36s7.064.76s1.055.47s5.49
Aug 23 8:00am3.96s16.675.30s0.947.23s2.49
Aug 23 7:00am4.95s16.1612.1s0.415.31s3.20
Aug 23 6:00am5.72s12.235.69s0.885.11s5.68
Aug 23 5:00am4.37s20.824.97s1.014.58s3.71
Aug 23 4:00am4.39s10.714.63s1.084.62s3.68

Last check 4m ago

69 records in view · all times America/New_York

The apps we are signed into

We buy the same paid plans you can buy. If a sign-in goes stale, we skip that app for the round and say so here, rather than publishing a broken number.

Claude Code sign-in

Signed in · claude.ai

claude.ai · max · v2.1.241 (Claude Code)

Codex sign-in

Signed in · ChatGPT

chatgpt · vcodex-cli 0.147.0

Grok Build sign-in

Signed in · grok.com

session · vgrok 1.0.5 (5115b46bc9) [stable]

No API keys

None set — we test the paid apps only

An API key in the environment would quietly move billing to the pay-as-you-go API, and we would stop measuring the paid app people actually use. Our runner strips those keys, and stops on purpose if one is set.

Test kit v0.2 · win32. App versions change often — a jump can move the numbers.

What the last round did

Every model the last speed round touched, in order. Recorded means we saved a measurement. Skipped means it was not that model’s turn, or the app had already hit a rate limit.

Show every target from the last round (13)
ping · codex
saved · ok · ok · 5.94s
ping · grok
saved · ok · ok · 4.63s
ping · claude
saved · ok · ok · 3.21s
speed · claude/fable
saved · failed · rate limit warning · 7.82s
speed · codex/luna
saved · ok · ok · 9.87s
speed · codex/terra
saved · ok · ok · 9.54s
speed · codex/sol
saved · ok · ok · 10.8s
speed · codex/gpt-5.5
saved · ok · ok · 10.3s
speed · grok/grok-4.6
saved · ok · ok · 5.95s
speed · grok/grok-4.5
saved · ok · ok · 6.53s
speed · claude/haiku
saved · ok · ok · 6.00s
speed · claude/sonnet
saved · ok · ok · 4.50s
speed · claude/opus
saved · failed · overloaded · 3m 15s

Read next