Real tests of the AI coding tools you pay for.

Claude Code, Codex, and Grok Build — timed on the same paid plans you buy. How fast they write, and whether the answer is right. For people on a subscription, not an API key.

See today's report How we test

SpeedWall tok/s

Live33m ago—

Grok 4.641.59
Opus 531.83
Fable 530.30
GPT-5.6 Sol22.54

Measured: 10:02 PM
Timezone: ET
Test kit: v6

Models measured: 10; Every round
Rounds a day: 24; One every hour
Clean rounds: 99%; 318 of 320 in 24 h
Rate limits: 0; In the last 24 h

Highlights

Today, in two numbers

The same two questions we ask every round, for each lab's leading model: how fast it wrote, and how often it finished a clean round.

Speed

Middle wall tokens per second, last 24 h · higher is faster

Grok 4.632/32 clean41.59
Opus 532/32 clean31.83
Fable 532/32 clean30.30
GPT-5.6 Sol32/32 clean22.54

041.59 tok/s

4 of 10 models

Other models

Finished the round

Clean rounds out of rounds tried, last 24 h · higher is better

100%

All 4 tied at 32/32 clean. Nothing to rank in this window.

4 of 10 models

Lab status

Both boards cover the last 24 hours. Each board has its own scale · test kit v6. 4 of 10 pinned models shown · the rest live on Other models.

Speed

Speed, ranked

The 4 models we lead with, ranked on the middle wall tok/s of their clean rounds in the last 24 hours. Longer bar is faster.

Other models

Tokens per second, start to finish. Higher is faster.

The 4 models on the home boards, ranked on median wall tokens per second over the last 24 hours.
Rank	Model	Wall tok/s	Clean	Rounds
1	Grok 4.6xAI · effort xhigh	41.59	100%	32/32
2	Opus 5Anthropic · effort high	31.83	100%	32/32
3	Fable 5Anthropic · effort high	30.30	100%	32/32
4	GPT-5.6 SolOpenAI · effort high	22.54	100%	32/32

Window 24 hours · 4 of 4 models scored · n = 128 rounds tried · test kit v6. Day-over-day change appears here once the lab has two full days of rounds in the log. 4 of 10 pinned models shown · the rest live on Other models.

24 hours

The last 24 hours, round by round

Each line is one model's middle speed over an hour of rounds. The faint dots behind it are the single rounds, so the swings stay in plain sight.

Open today's report

Wall tokens per second · higher is faster

nothing recorded before 12:00am · the plot starts there

tok/s020406080

12am2am4am6am8am10am12pm2pm4pm6pm8pmnow

Aug 20

Line: the middle wall tok/s of each 4 hours of testing — up to 4 rounds. Dots: those single rounds. Nothing is filled in, so a stretch with no clean round breaks the line, or ends it early.

Grok 4.6median41.5923/24 cleanOpus 5median31.8323/24 cleanFable 5median30.3023/24 cleanGPT-5.6 Solmedian22.5423/24 clean

Full record — every round, every number

Wall tokens per second for every round in the last 24 hours. A dash means that round had no clean result.
Round	Grok 4.6	Opus 5	Fable 5	GPT-5.6 Sol
Aug 20 10:00pm	41.21	32.36	28.66	15.38
Aug 20 9:00pm	40.70	28.77	30.49	22.34
Aug 20 8:00pm	39.51	35.12	31.81	26.26
Aug 20 7:00pm	49.41	30.27	32.04	23.67
Aug 20 6:00pm	38.32	33.52	31.43	26.76
Aug 20 5:00pm	49.19	41.54	30.11	23.22
Aug 20 4:00pm	43.28	32.95	29.64	17.36
Aug 20 3:00pm	47.80	36.57	32.90	24.96
Aug 20 2:00pm	37.03	33.50	27.28	24.98
Aug 20 1:00pm	48.43	34.58	32.47	16.95
Aug 20 12:00pm	50.03	31.22	31.66	21.82
Aug 20 11:00am	34.39	33.35	26.64	19.65
Aug 20 10:00am	45.29	28.59	31.80	22.56
Aug 20 9:00am	53.94	26.77	30.04	22.23
Aug 20 8:00am	43.63	32.08	32.10	19.56
Aug 20 7:00am	39.61	33.76	29.87	26.41
Aug 20 6:00am	5.55	33.45	24.65	24.46
Aug 20 5:00am	40.48	31.57	34.60	21.38
Aug 20 4:00am	45.86	31.24	29.75	22.57
Aug 20 3:00am	32.50	31.94	23.86	24.68
Aug 20 2:00am	44.12	29.39	32.65	23.79
Aug 20 1:00am	41.27	25.98	29.01	20.34
Aug 20 12:00am	17.67	33.57	18.98	22.52
Aug 19 11:00pm	—	—	—	—

Window 24 hours · one round every 60 minutes · 24 rounds on the clock · n = 92 clean readings drawn · test kit v6. The lab recorded nothing in the first 1 hours of this window, so the plot starts at 12:00am rather than drawing empty space. Those rounds are still in the full record above, as dashes. 4 of 10 pinned models shown · the rest live on Other models.

We also ping each app every round to check it answers at all. Those checks, round by round, live on Lab status.

Method

How we get these numbers

We run the real paid apps on our own lab machine, not the hidden APIs. Every number traces back to a saved run.

Open How we test

Cadence: Every hour
Rounds a day: 24
Test kit: v6
Window: 24hours

Test kit v6: Old rows stay in Logs. They do not mix into these boards.; Changelog
Paid apps: We run Claude Code, Codex, and Grok Build ourselves. Not the hidden APIs.; Status
Graded answers: A round counts only when the answer is right. Not “it printed text.”; How we test
Saved proof: A public number traces back to a saved run in our archive.; Logs

Last round 35m ago · 483 results saved in the last 24 hours, 5 failed · Lab status.