TL;DR: Grok 4.6 scores the same as the model I pay for at about a sixth of the cost. I am not moving, and not because of the training toggle. Every paid plan has one and everyone should switch it off. The difference is in what you cannot switch off: the licence xAI takes over your content, and its refusal to say how long it keeps anything.
A chart went round this week. Two models, same score on a coding benchmark, and one of them costs about a sixth of what the other does.
I worked out what it would save me over a year, and then did nothing.
The model really is that good
No point pretending otherwise. Grok 4.6 landed on 12 August and on Cursor's benchmark it ties the leader, at a sixth of the cost. It is not a bigger model. It is more training on the same base as Grok 4.5, and it gets there in roughly half the steps.
The price is not a promotion either. SpaceX put 86% of its entire capital budget into AI last quarter, at a company that also launches rockets, and Musk wants 10 GW of compute running by the end of next year.
Worth pausing on that number. For comparison, New York City at full summer peak draws about 10 GW. Britain's whole electricity demand averages around 36 GW (320 TWh a year, spread over the hours in it).
Build that much capacity that fast and the price per task falls. That is the whole mechanism.
The training toggle is the easy half
Everyone knows this part, or should. Every paid consumer plan trains on your conversations unless you turn it off. Claude, ChatGPT, Gemini, Grok. All four, same shape.
In Claude it is Settings, then Privacy, then "Help improve our AI models". Leave it on and Anthropic keeps your data five years. Turn it off and it is thirty days. If you have never opened that screen, go and do it now, it takes ten seconds.
But that is not a reason to pick one vendor over another, because they all give you the same switch. It is hygiene, not a differentiator.
What you cannot switch off
And highly likely you’d prefer to.