I’ve entered a fresh, new, valid Anthropic API key. I’ve toggled ON to use the key. I’ve selected an Anthropic model in my chat window. I have checked and double checked I have set a budget for my Anthropic API keys. My other keys in other apps I’ve deployed work great. But in Cursor IDE, I absolutely can not prompt any Claude model. I’ve tried adding the model names via “add custom model”. I even downgraded from the beta, just to see if it was a beta issue. DOES NOT WORK.
Steps to Reproduce
Follow your own steps in the Cursor Settings > Models > API Keys for connecting an Anthropic key.
Expected Behavior
I should be able to plug in an API key, slide the slider, and it should just work. Also, what happened to your Verify button? You used to be able to verify your key is active and working.
Opus 4.5, Opus 4.5 thinking, Sonnet 4.5, Sonnet 4.5 thinking
For AI issues: add Request ID with privacy disabled
Request ID: 2642fe59-2122-4a4a-b882-d8655ebf46e2
{“error”:57,“details”:{“title”:“Provider Error”,“detail”:“We’re having trouble connecting to the model provider. This might be temporary - please try again in a moment.”,“isRetryable”:true,“additionalInfo”:{},“buttons”:,“planChoices”:},“isExpected”:true}
Same problem here. Never encountered this issue before in the couple months i’ve been using Opus 4.5 thinking. Thing’s I’ve tried so far, none of which worked:
Switching to Opus 4.5 (not thinking)
Switching to Sonnet 4.5
Adding an Anthropic API key in Cursor’s Models settings and retrying 1 and 2
Switching to non-Anthropic models resolves the issue, but doesn’t help me as I’m halfway done building out a major feature on our app and have found Opus 4.5 to be more reliable than other models. So dead in the water right now. Going to give Claude Code a try and check back later to see if this is resolved.
Yeah, it’s super frustrating. If I used Claude Code I’d get (reportedly) $1000 worth of API requests for $200/mo, but then I’d have to use Claude Code (please don’t make me, Cursor).
I’m currently spending anywhere up to $800 USD/mo on Cursor. All I’m trying to do is bring that down by maybe $100 by accessing the Anthropic API directly - and even that might be optimistic. My guess is that, because it’s not as profitable fix for Cursor, they don’t really care whether or not we have access to the API’s directly. It effectively routes my money away from Cursor and directly to the models we’re using.
It really makes me wonder about the profit model of Cursor in general. I like having the flexibility to move to use other models if a new, better (non-Anthropic) model comes out tomorrow. But is it worth whatever cut (completely opaque btw) Cursor is taking just for that privilege? Unclear.
If they actually offered me a $40/mo BYOK (bring your own keys) plan, I’d consider it. It’s really unclear with the current interface what is actually the most economical for me the user, and when you’re a power user dropping this kind of money, it doesn’t make me feel good about staying with Cursor.
But fwiw I just found a workaround you might want to try if you haven’t already: start a new agent thread. Both Opus models are working for me in the new thread. And I was able to get it up to speed by pointing it to the plan document that the prior agent was working off of, and then manually copy/pasting the pending todo list items from the old agent to the new one. Good luck!
Check your OpenAI settings: Cursor Settings > Models. Make sure there’s no OpenAI Base URL override or OpenAI API Key set. Those can conflict with Anthropic requests. If you see them, clear them.
Start a new chat: @HellyFurtado mentioned that a fresh chat helped. Try starting a new one.
Verify button: It was removed because it often caused an error, so when you add your API key, you don’t need verification.
If that doesn’t help:
Network Diagnostics: Cursor Settings > Network > Run Diagnostics
DevTools: Help > Toggle Developer Tools > Console tab. Send the errors you see when you try the request.
Let me know if starting a new thread or disabling HTTP/2 helped. If not, I’ll share the diagnostics results with the team for investigation.