My first deep dive with Claude Code was in Jan of this year. I was working on rebuilding an old city builder game from the the 00s and was trying to parse the original images from the game files. There was plenty if info on the structure online so I said I'd let Claude give it a go. I was running one of the sonnets I think.
After about 5 prompts I had a web interface that I could drag the files into and it presented me a grid of all the images and I was able to export more common sprite sheets. Whole thing took maybe 10-15min and my mind was blown.
I have tried to do the same recently with the newer models and we're still not writing code by the time the older models had finished their work. Something has definetly changed in relation to how much pre work the newer models are doing and maybe it's by design for more consistant results.
Not commenting on speed, but it seems models and their harnesses have generally gotten much worse over time. My suspicion is that the "Frontier" labs really dont have a strong handle on good quality evals that equal expectations of their users, so they just churn out new models for marketing to sell the shit out of.
Recently cancelled my Codex subscription because I cannot stand 5.6-sol/terra/luna. On top of that, Codex the harness is just so dammed buggy in VS Code.
I just discovered https://tinfoil.sh, which is fully private and attestable AI which works amazing with Kilo Code. Its costing me at API prices but for cheaper to run models which so far feel more consistent then what I get from frontier models.
As someone working in this space: unless they're a very big player with resources to create custom hardware, afaik there's no way to actually prevent the GPU host from seeing the request content. There's no "secure enclave" that sits between a CPU and the GPU, the decrypted payload must necessarily hit system RAM, as would the output. They are, at best, running in TEE VMs which are also pretty horribly broken and don't provide the security that they say they do.
> I have tried to do the same recently with the newer models and we're still not writing code by the time the older models had finished their work
Claude used to run at ~150tps in January. Now it runs at 50tps if you're lucky. That explains most of the speed issues without even getting into technical details about the new models.
Fable is quite good, but i also noticed that the later opus models don't seem quite as effective as before.
We break down our tasks pretty granularly before they get picked up by a model, and for that workflow we've found that sonnet 4 and opus 4 are still quite effective, and debatably more effective than the 5s
I do too. In the past couple of months it's just doing too much. Before it was prompt -> model writes code -> done, I check the work. Now it's prompt -> model seems to read the entire codebase to 'get the full picture' -> takes forever to write code -> runs a whole bunch of things to verify the result. It seems like they have optimized the harness more and more for people who don't know what they are doing or what they want. I know what I want. I mostly want to 10x my typing, I already know what I want and what files to edit and how to verify the result.
My first deep dive with Claude Code was in Jan of this year. I was working on rebuilding an old city builder game from the the 00s and was trying to parse the original images from the game files. There was plenty if info on the structure online so I said I'd let Claude give it a go. I was running one of the sonnets I think.
After about 5 prompts I had a web interface that I could drag the files into and it presented me a grid of all the images and I was able to export more common sprite sheets. Whole thing took maybe 10-15min and my mind was blown.
I have tried to do the same recently with the newer models and we're still not writing code by the time the older models had finished their work. Something has definetly changed in relation to how much pre work the newer models are doing and maybe it's by design for more consistant results.
Not commenting on speed, but it seems models and their harnesses have generally gotten much worse over time. My suspicion is that the "Frontier" labs really dont have a strong handle on good quality evals that equal expectations of their users, so they just churn out new models for marketing to sell the shit out of.
Recently cancelled my Codex subscription because I cannot stand 5.6-sol/terra/luna. On top of that, Codex the harness is just so dammed buggy in VS Code.
I just discovered https://tinfoil.sh, which is fully private and attestable AI which works amazing with Kilo Code. Its costing me at API prices but for cheaper to run models which so far feel more consistent then what I get from frontier models.
As someone working in this space: unless they're a very big player with resources to create custom hardware, afaik there's no way to actually prevent the GPU host from seeing the request content. There's no "secure enclave" that sits between a CPU and the GPU, the decrypted payload must necessarily hit system RAM, as would the output. They are, at best, running in TEE VMs which are also pretty horribly broken and don't provide the security that they say they do.
Tinfoil is almost certainly lying to you.
[dead]
> I have tried to do the same recently with the newer models and we're still not writing code by the time the older models had finished their work
Claude used to run at ~150tps in January. Now it runs at 50tps if you're lucky. That explains most of the speed issues without even getting into technical details about the new models.
Oh that's interesting. When did that happen? a 3x reduction in speed would probably explain it
It happened as their user base exploded and their infrastructure could not keep up with all the new demand.
I miss humans.
Fable is quite good, but i also noticed that the later opus models don't seem quite as effective as before.
We break down our tasks pretty granularly before they get picked up by a model, and for that workflow we've found that sonnet 4 and opus 4 are still quite effective, and debatably more effective than the 5s
for reference, we use n-dx (https://n-dx.dev) for our workflow
I do too. In the past couple of months it's just doing too much. Before it was prompt -> model writes code -> done, I check the work. Now it's prompt -> model seems to read the entire codebase to 'get the full picture' -> takes forever to write code -> runs a whole bunch of things to verify the result. It seems like they have optimized the harness more and more for people who don't know what they are doing or what they want. I know what I want. I mostly want to 10x my typing, I already know what I want and what files to edit and how to verify the result.
Yes, author here. I was thinking about it more after the post.
I wonder if the harness has a bigger share of blame here than the model itself.
I seem to like using Opus 5 in the browsers a lot more... I.e. it was very helpful in fine-tuning my site after it went down due to HN traffic.
Site is broken: "Error establishing a database connection"
Fixed now, I did use Opus 5 to help me fine tune it :)
Will probably follow up with another blog post about that soon.
https://web.archive.org/web/20260826063455/https://alexkras....
Thank you!
If only old claude was here to fix it.
It did help eventually :)
Seems the site is down
Fixed now, thank you.
it won me over at 4.6 and lost at 4.7 and the image 'How multiagent turf war runs ended' [0] might explain the shift in the behaviour
[0] https://www.anthropic.com/research/multiagent-systems
I too keep setting Opus 4.6 manually in my daily work (when I am not using older Sonnet).
We all do.
[flagged]