Comment by EB66
2 days ago
> I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.
I've assumed the same as well.
I also assume that many of the companies developing these models engage in benchmaxxing.
At my company we've developed our own internal benchmarks for evaluating LLM models as they become available. The benchmarks are tailored to our particular use cases but the utility and knowledge our benchmarks assess is still fairly universally applicable. I see wide differences between what our internal benchmarks report and what the major benchmarks do.
There was a whole lot of fanfare about how amazing GLM 5.2 was when it was released, but it was pure rubbish on our internal benchmark -- far behind OpenAI, Anthropic, Gemini, DeepSeek, etc. I don't know how to reconcile the fact that GLM 5.2 performed very well on some of the major public benchmarks, but consistently performs so poorly on ours. OpenAI models tend to dominate our internal benchmarks.
It actually is showing in public benchmark if you know how to look for it. For example, in Terminal-Bench 2.1, GLM 5.2 received 78%, while GPT 5.6 Sol received 88%.
Then Terminal-Bench 3.0 came out (where the questions are new), and GPT 5.6 Sol received 34.6%, while GLM 5.2 dropped to a whopping 4.6%.
> OpenAI models tend to dominate our internal benchmarks.
That's odd, since Fable seems to be the leader for the industry. Not cost-effective, but if Anthropic models get dominated by OpenAI in your internal benchmarks, this calls their validity into question. Separately, see the jagged frontier effect. [1]
I've been using GLM 5.2 at my day job (mostly Rust backend work ATM). Nothing that blows away the models from OpenAI and Anthropic, but solidly good enough to get it done. A lot of people have experienced this and the fact that an open weights model can do so is where most of the excitement comes from. Optimizing for benchmarks can only get you so far, and people are quick to criticize models that fall into it (like DeepSeek Pro V4 recently).
[1] https://mitsloan.mit.edu/ideas-made-to-matter/working-defini...
This resonates. You need to tailor evaluation to your own use cases in order to know if you should adopt the next thing or not.