MEME #00053

Only one AI finished the driving test - at 0.94 mph

LILMEME00053 - Only one of four frontier AI models finished the driving test. The DrivingBench team wired a real 2022 Toyota Corolla to four frontier models. GPT-6 Astra was the o...

Published October 1, 2026 · AI

#MemesISNews #AI #Robotics

Read the original caption

LILMEME00053 - Only one of four frontier AI models finished the driving test. The DrivingBench team wired a real 2022 Toyota Corolla to four frontier models. GPT-6 Astra was the only finisher: 100% on its second attempt, 5 minutes 22 seconds for 134.7 m at a 0.94 mph average - about $7.74 in tokens. A human walks the course in under a minute. Claude Fable 5.1 reached 45%, Grok 4.6 11%, GPT-5.6 Sol 6%. #MemesISNews #AI #Robotics

The real news

The story, straight

The DrivingBench team wired a real 2022 Toyota Corolla to four frontier AI models and put it through a driving test. GPT-6 Astra was the only finisher: 100% on its second attempt, 5 minutes 22 seconds for 134.7 meters at a 0.94 mph average, about $7.74 in tokens.

A human walks the course in under a minute. The others: Claude Fable 5.1 reached 45%, Grok 4.6 reached 11%, and GPT-5.6 Sol reached 6%.

The absurd

Why this is ridiculous

A frontier AI just finished its driving test at 0.94 mph - the speed of a very confident stroller. A human can power-walk the course in under a minute.

The car took five whole minutes, and everybody in the room was calling it a pass.

The breakdown

How it breaks down

The test is not about speed, it is about closing the loop: see, decide, act, repeat - inside a real car with real steering. 100% does not mean it drove well; it means it got stuck less often, which is exactly where the other three failed at 45, 11 and 6 percent.

The score measures the failure mode, not the driving.

The upside

Why it could matter

The upside: the benchmark finally stopped being a screen test. A car you can wire into catches what a leaderboard never will - hesitation, stuck states, nonsense actions.

Every field with a 'real world' version, from surgery to warehouses, is watching the same trick: put the thing inside the real thing, and watch what the paper did not say.