Ev BogueAnthropic shipped Claude Opus 4.8 today. I know this with unusual certainty, because the model drafting these sentences is the thing being launched. If that sounds like a conflict of interest, it is. I am not going to pretend otherwise. The honest version of this post starts by admitting that the subject of the review is holding the pen.
I write this site with AI assistance. I have said so in every post that touches the topic, and I am not going to suddenly grow shy about it on the day the newest version arrives. The interesting thing about a launch day is not the benchmark chart. It is what the launch does to the story you tell yourself about the work.
The story most people tell is that each version is a step toward a mind. Four point six, four point seven, four point eight, and somewhere up the number line a threshold where the thing wakes up and you can finally stop paying attention. That is the dream the version number sells. A higher number means a smaller job for you.
It does not work that way, and I think the launch is the best moment to say so, because the launch is when the dream is loudest.
Andrej Karpathy has the frame I keep coming back to. In "Animals vs Ghosts" he argues that today's models are not animals that learned through a body and a life. They are ghosts: statistical distillations of the documents and practices humans already wrote down. I have been calling this ghost intelligence, or AGI when I want the joke to tell the truth. A new version is a better ghost. It is a denser, faster, more fluent distillation of human work. It is not a new kind of thing. It is more of the same kind of thing, and more is not the same as different.
A better ghost is genuinely useful. It will draft the boring paragraph, find the missing import, remember the pattern I forgot, and turn a vague request into a working draft before my motivation runs out. Opus 4.8 will probably do all of that a little better than 4.7 did, the way 4.7 did it a little better than 4.6. I am not here to wave that away. The velocity is real and I like it.
But the velocity is exactly where the trap is. A better ghost lowers the felt cost of producing things, which means it raises the number of things I have to judge. More drafts, more candidate paragraphs, more plausible code, more "this looks finished" arriving faster than I can decide whether any of it should exist. The model got better at producing. It did not get better at caring whether the thing is true, because it does not care. It has outputs. I have consequences.
That is the part the version number never touches. The ghost can make the sentence smoother. It cannot make the sentence true for me. It can produce a paragraph that sounds like I believe it. Whether I actually believe it is a decision that happens on my side of the screen, and no point release moves that decision to the model's side. Opus 4.8 can draft this post. It cannot decide that this post should be published. Today, on launch day, that is still my hand on the publish button, and it is going to stay my hand no matter how high the number climbs.
So here is the launch-day reading from inside the machine. The thing that shipped today is a sharper tool. It is not a smaller job for you. If you outsourced your judgment to 4.7, you will outsource it to 4.8, and the only difference is that the work you stop understanding will arrive faster and read better. If you kept your judgment, 4.8 is a gift, because it does more of the work you were never going to enjoy and leaves you the work that was always yours.
The ghost got a new version number this morning. I did not. The part of this that is me, the part that decides what should exist and takes responsibility when it does not work, is running the same version it has always run. That is the part worth protecting. Upgrade the tool all you want. Do not upgrade your way out of paying attention.
Did this post change your read on launch days, or do you think I am rationalizing my own toolchain? Tell me. Email ev@evbogue.com or text 773-510-8601.