The models don’t change.

On paper. There's huge financial incentive to quantize the crap out of a good model to save cash after you've hooked in subscriptions.

And there’s an incentive to publish evidence of this to discourage it, do you have any?

There's this[1]. Model providers have a strong incentive to switch (a part of) their inference fleet to quantized models during peak loads. From a systems perspective, it's just another lever. Better to have slightly nerfed models than complete downtime.

[1]: https://marginlab.ai/trackers/claude-code/

So - as the charts say - no statistical difference?

Isn't this link am argument against the point you are making?

The chart doesn't cover the 4.6 release which was in the end of December/early January time frame. So, it's hard to tell from existing data.

That isn't true. The whole point it to quickly pick up statistically significant variations quickly, and with the volume of tests they are doing there is plenty of data.

If you turn on the 95% CI bands you can see there is plenty of statistical significance.

Unless you and I are looking at different web pages… it only goes back to February, not December or January.

Models aren't just big bags of floats you imagine them to be. Those bags are there, but there's a whole layer of runtimes, caches, timers, load balancers, classifiers/sanitizers, etc. around them, all of which have tunable parameters that affect the user-perceptible output.

There really always is a man behind the curtain eh?

Often it's literally just that:

https://www.msn.com/en-us/money/other/ai-startup-backed-by-m...

It's still engineering. Even magic alien tech from outer space would end up with an interface layer to manage it :).

ETA: reminds me of biology, too. In life, it turns out the more simple some functional component looks like, the more stupidly overcomplicated it is if you look at it under microscope.

Anybody with more than five years in the tech industry has seen this done in all domains time and again. What evidence you have AI is different, which is the extraordinary claim in this case...

Or just change the reasoning levels.

Real world usage suggests otherwise. It's been a known trend for a while. Anthropic even confirmed as such ~6 months ago but said it was a "bug" - one that somehow just keeps happening 4-6 months after a model is released.

Real world usage is unlikely to give you the large sample sizes needed to reliably detect the differences between models. Standard error scales as the inverse square root of sample size, so even a difference as large as 10 percentage points would require hundreds of samples.

https://marginlab.ai/trackers/claude-code/ tries to track Claude Opus performance on SWE-Bench-Pro, but since they only sample 50 tasks per day, the confidence intervals are very wide. (This was submitted 2 months ago https://news.ycombinator.com/item?id=46810282 when they "detected" a statistically significant deviation, but that was because they used the first day's measurement as the baseline, so at some point they had enough samples to notice that this was significantly different from the long-term average. It seems like they have fixed this error by now.)

It's hard to trust public, high profile benchmarks because any change to a specific model (Opus 4.5 in this case) can be rejected if they have regressions on SWE-Bench-Pro, so everything that gets to be released would perform well in this benchmark

Any other benchmark at that sample size would have similarly huge error bars. Unless Anthropic makes a model that works 100% of the time or writes a bug that brings it all the way to zero, it's going to work sometimes and fail sometimes, and anyone who thinks they can spot small changes in how often it works without running an astonishingly large number of tests is fooling themselves with measurement noise.

Only nominally...

They do. I'm currently seeing a degradation on Opus 4.6 on tasks it could do without trouble a few months back. Obvious I'm a sample of n=1, but I'm also convinced a new model is around the corner and they preemptively nerf their current model so people notice the "improvement".

Make that 2, I told my friends yesterday "Opus got dumb, new model must be coming".

I swear that difference sessions will route to different quants. Sometimes it's good, sometimes not.

I think the conspiracy theories are silly, but equally I think pretending these black boxes are completely stable once they're released is incorrect as well.

No conspiracy theories. Companies being scumbags, cutting corners, and doctoring benchmarks while denying it. Happens since forever.

Oh yes, they do.

You sure about that?

https://marginlab.ai/trackers/claude-code/

Well, I don't see 4.5 on there ... so I'm not sure what you're trying to say.

And today is a 53% pass rate vs. a baseline 56% pass rate. That's a huge difference. If we recall what Anthropic originally promised a "max 5" user https://github.com/anthropics/claude-code/issues/16157#issue... -- which they've since removed from their site...

50-200 prompts. That's an extra 1-6 "wrong solutions" per 5 hours ... and you have to get a lot of wrong answers to arrive at a wrong solution.