Using Fable, pretty much every request hit some gate they had for no discernible reason. These provider-level rejections should be incorporated into benchmarks as 0s on the tasks since that's the experience you'll actually get using the model.
I have heard this from a bunch of folks, but that was not my experience. For the couple days I was able to use it, I didn't hit a single gate, and I was using it pretty extensively (but not for anything security-related).
Seems like the model is incredibly inefficient at max reasoning, and even at high/xhigh it uses far more tokens than other models, including Gemini 3.5 Flash, GLM 5.2 and so on. GPT 5.5's efficiency in tokens is still unmatched.
Half of the data is missing and the rest is inconsistent between different graphs and sections. Is the benchmark having Sonnet 5 generate the page and seeing how many hallucinations it has?
It starts with NVIDIA artificially and slowly releasing its tech. If the GPUs were cheaper, we would have better models by many other companies, and competition would take care of these greedy tactics.
I used sonnet five today to evaluate work I’m doing on an experimental programming language with an interesting concurrency model.
I asked it to try to figure out why one of the examples wasn’t working.
It read the implementation of the compiler and the runtime, found the bug, fixed it, fixed the example and the only thing I had to do manually is suggest a less silly name for a particular function.
See also: https://cursor.com/cursorbench
Cost per task data is only available for max effort though, might just be very inefficient at that effort level.
A release just to have a headline while Fable situation is getting resolved.
I asked it to try to figure out why one of the examples wasn’t working.
It read the implementation of the compiler and the runtime, found the bug, fixed it, fixed the example and the only thing I had to do manually is suggest a less silly name for a particular function.
I would use sonnet 5 for coding … seems alright!