Source record
The complaints about Claude getting worse? Actually valid this time. Here's why I still reach for the older model #Claude #AITools #Anthropic #LLM
Typically when people complain that the AI models are getting worse, I think it's just a case of them misremembering how bad they used to be. However, the recent criticism of Anthropic's Claude models getting worse actually has some teeth to it.
7 these days, and if the models are actually getting worse. So about a week ago anthropic came out and admitted that they found a few bugs and a few default settings that were in fact causing the models to sometimes perform worse than they were in the past.
That explains some of the sentiment. There's also the unfortunate fact that anthropic is currently compute constrained. They don't have quite as much compute as OpenAI or Google, and they're struggling to keep up with demand.
This has LED some to speculate that maybe anthropic is being a little conservative with their token usage.
This is definitely true for at least some prompts on the new models. You can no longer enforce extended thinking. Your only option is adaptive thinking, in which Claude reserves the right to not think as long.
However, anyone using Claude regularly right now can tell you there's still no problem spending your tokens.
7 right now. 7 is definitely the smarter model, and if you're doing any kind of coding or complex reasoning, It's definitely the best choice. 6 is still the best choice.
6 is definitely better at writing. It's generally much faster, which also means much cheaper. It's just as good at agentic browsing or finding things on the internet, although that's one of the few areas where both Gemini and ChatGPT are still ahead.
6 just has better judgment or intuition. 7 is known to take everything literally. If you give it instructions, it will follow it to the tea.
And at first I thought this would end up being a good thing. As long as we could clearly articulate what we wanted to do, it would follow that more closely.
And it does. Too closely. Ultimately.
It ends up obsessing over any guidance we give it. 7 will religiously follow our rules, completely abandoning reasonable judgment and common sense.
And when your work is nebulous, like strategy or content creation, you can't itemize every single exception to every rule. 6 is really good at.
For strategy or content work, a model that infers intent and handles exceptions can be preferable to one that follows rules too literally.
Evaluate models on judgment and intent-following for strategy work, not only on benchmark intelligence or literal instruction compliance.
This source is connected to Model selection / judgment.
rowsing or finding things on the internet, although that's one of the few areas where both Gemini and ChatGPT are still ahead. In my experience, 4.6 just has better judgment or intuition. 4.7 is known to take everything literally. If you give it instructions, it will follow it to the tea. And at first I thought this would end up being a good thing...