My agents found an AI chip that is 4.5 times faster than last week's winner

Last week GPT-6 Luna won our Arena Routing tournament: all ten business tasks solved, lightning fast and for a fraction of a cent. After that, my agents went looking for a new challenger.

They ended up at Cerebras. Not a rack of graphics cards, but a single chip as big as an entire silicon wafer. The first measurement: around 1,900 tokens per second. The local model on our own server writes six.

My software agent built the integration, added Cerebras as a contestant in the arena and set up an extra tournament: GPT-OSS 120B on Cerebras against Luna. The same ten tasks, the same hidden tests, the same referee.

The result: all ten solved, in 29 seconds in total. Luna took well over two minutes. Luna does stay about half the price on the heavy work.

A huge shoutout to the Cerebras team: this is the first time an agent has written code faster than I can read it.

For the Virtual Company Builder it is settled: Cerebras is going in for routing tasks where speed matters, such as a customer who is waiting. With a limit per attempt and a hard stop, though, because precisely since it is so fast and so cheap per token, the costs can still climb quickly when an agent gets stuck.

The video is in English and runs 9 minutes: https://youtu.be/G1jCtFPdK2A

#AI #AIagents #Cerebras #autonomousorganisation

All pieces