12.8 to 62.7
That is the jump in DeepSeek's own agentic-coding score between the V4 preview and V4-Pro, which came out of preview last Thursday; terminal-operation scores went from 72.1 to 87.9. Vendor numbers, so discount them accordingly, but the direction is not in doubt. This generation of models is not being trained to talk to you. It is being trained to finish a job, one step at a time, unattended.
Then there is the pricing. From 17 August, DeepSeek charges peak and off-peak rates, double during Beijing office hours, half the rest of the time. Compute is starting to be sold the way electricity is: work at night, pay half.
From 17 August, DeepSeek charges peak and off-peak rates, double during Beijing office hours, half the rest of the time.
I use AI agents daily in my own work, and the practical shift is this, the valuable skill moves from asking to checking. A good prompt is now the entry fee. Spotting which of its ten steps went wrong, and being willing to send it back, is the hard part.
Most of what schools teach is still how to hold a conversation with AI and get it to write something.
The colleague our children will work alongside does not keep office hours. It works at 3am, at half price.
