Asked it because the answers were obviously not as it used to, and to my surprise, I can see why now…
4 Likes
I think it should show a warning when redirected to another model (due to outage, or anything else). Also I hope it doesn’t count toward the monthly count.
2 Likes
Hi @m955
Anthropic had a few incidents today, so you were automatically switched to another model:
Rerouting models without user knowing isn’t very transparent and, in my opinion, is misleading and untrustworthy. I would rather have failed requests to Anthropic and be certain that the model I’m using is actually Anthropic. Sometimes I really feel like I’m not getting what I’m paying for.
Sonnet seems to perform poorly at times, even when it’s up, going in endless loops without solving my problem. And when I switch with my API key, it’s doing it in 0-shot. This is really strange, and it’s disappointing to see that there’s logic for this in Cursor, so that could be true.
4 Likes
Didn’t you get any notifications about the model change? It should be there, and it’s up to you to decide whether to wait for the model you need to work or choose another one.
Did you try with o1-mini or Gemini? I don’t think it’s the model but the complexity of the problem + changes in prompting methods could also help
I am talking here from a coding session 1 week ago where i used normally my Fast-Requests and Anthropic had no downtime and it did not solve my problem after 5+ prompts. Then i added the same files for context and did the same prompt, and apparently, with my Anthropic API key, it solved it 0-shot. This is somehow weird if you ask me.
3 Likes
They updated to the new Sonnet version yesterday, maybe that was it then.
We never switch to GPT-3.5 when anthropic has downtime. We switch to GPT-4o only when anthropic has downtime. The downtime was momentary enough yesterday that we did not show the error message to users, but we’ll be sure to do this next time.
The model will tell you it is GPT-3.5 because the switch is done in a way that re-uses our sonnet prompts. So when GPT-4o sees a claude-3.5 prompt, it sees 3.5 and thinks GPT-3.5
In general we’ve been very frustrated with Anthropic’s reliability and hope as they scale they will have fewer incidents. It’s most frustrating when they have downtime that isn’t reported on the status page so users think we’re switching models just to save money.
3 Likes
Same for me. For three months, I have been a pro subscriber. I used cursor to fix a script with Sonnet 3.5’s long context (200k) chat and ended up in a vicious cycle: 7 prompts with nothing only made things worse. Then I tried the same prompt with the same context using Cline, and it resolved it in two shots. I am considering switching back to VScode and canceling my pro subscription. I believe that Cursor is over-advertised. The only drawback of Cline is that it consumes a large number of tokens, but it is worthwhile because it resolves the issue and saves the project.