[Solved] Add Claude 3 models

Yes so a limited amount of full context calls, maybe we can pay for them with our own API would be great. But there’s some work that goes into the prompt

1 Like

@truell20

Been experimenting with Opus and Sonnet using the Openrouter way. So far seems to be working. Anyone with more experience with this who can tell if some things won’t work or has encountered some weird issues?

EDIT: this seems a neat approach. Let the LLM providers (OpenAI, Anthropic, Google etc.) compete with capabilities and drive the price down and let the Cursor user pick the one serving you the best.

My observations and understanding:

  • Copilot++ is a separate model and not affected by the selected model
  • Codebase search uses Cursor indexing and retrieval (selected model does not have effect). Same with @docs etc.
  • Context is fed into selected model (Opus, sonnet etc…) for generation
  • Inline edit mode uses the selected model

My guesses:

  • Problems may arise if there are some hidden system prompting that might not be behave well with other than officially supported OpenAI models?
  • Problems may arise if Cursor uses some special format for diffs that is not working with Anthropic models?
1 Like

It’s suddenly working, did not update the IDE. Removed and added model type, API key and URL. Don’t know where the error could’ve come from.

I had it too, it worked the next day on its own :joy:
Here’s an article on how to set it up: Adding Claude 3 Opus (or any other model) to Cursor - Vincent Schmalbach

Any update for the Pro user?

1 Like

I would be interested in Claude 3 as well. Depending on the cost, I would be willing to pay more for a plan that included some of these models out there that are better for coding.
10 + more $ would be easy to pay for a better coding model. Maybe GPT will get there in time when V5 comes out but in the interim I think its worth it to us to have the option to pay more for whats considered the best coding LLM.
How easy this is to implement for the Cursor team, idk, but just wanted to put my vote out there.

Cursor team, LOVE what you have done and tell all my dev friends about you and I look forward to the future of this product.

1 Like

The next version of Cursor will have proper support for Claude 3 Opus, with usage-based pricing.

In the meantime, we will be releasing a backend fix tonight that should make it possible to use a small number of Claude 3 queries in the existing version of Cursor, at no extra cost.

8 Likes
  • Is there an ETA for the next version that will support the usage-based pricing?
  • Could you elaborate more on the usage-based pricing part? Will it be the same cost as the API, $15/1m input tokens $75/1m output tokens or did you manage to get some kind of special deal?
  • Will we be able to choose the context size? If not, what will it be set to.
  • Thoughts on supporting Haiku which just released Claude 3 Haiku: our fastest model yet \ Anthropic ?

That’s disappointing to be honest. Cody included it and they charge half price, Phind who charge the same, included it as well (500 per day).

I paid for year in advance, is it what I should expect, that basically in the future if better model cost more I won’t have access to them without paying for my own usage?

2 Likes

Thank you to all of you.
I don’t know when or how much additional fees will be added, but this is a very welcome update.

1 Like
  1. Next version should come in a couple of days.
  2. Not completely decided yet, but will likely be the same cost as the API, once you exceed the limit of daily included requests.
  3. Currently all requests are using 10k context, but we are working on a “long context mode”, to be released soon, specifically for Gemini 1.5 Pro but potentially also for Claude 3.
  4. We’ll probably add in Haiku too (it is much easier for us to add models now).

Right now, if you’re a Pro or Business user, you can add “claude-3-opus” as a custom model in the Settings page, and use 10 fast requests per day (unlimited slow, but the delay increases exponentially). The exact specifics of the pricing going forward is yet to be decided. Let me know if it doesn’t work for you.

14 Likes

i am not affiliated with cursor other than being their certified number 1 fan and enterprise customer, and i am not trying to say that your frustration is unjustified, @debian3.

i do want to give the cursor team a bit of a break here, by simply pointing to another example of how challenging it can be for SAAS providers to try to get the LLM pricing right (balancing their margins, user happiness, and the risk that they accidentally go into the red on inference costs because they picked the wrong per message fee). this is from last week:

poe (which for those who don’t know, is a much larger team than cursor that is part of the multibillion dollar company quora) literally had to >3x the “points” price of claude-3-opus after just a few days, because they miscalculated how expensive it would be:

this is hard. we are all stoked for claude 3, and i can’t wait to test it myself inside cursor. i want to commend cursor for taking a thoughtful and transparent approach to figuring out the pricing for their users.

(& let’s not forget cursor’s generosity to offer “at cost” usage with your own API key! i am naturally suspicious of SAAS companies and this feature was a big part of what made me feel like i could truly trust the cursor team.)

8 Likes

In terms of price , I don’t think the claude-3-opus would be worth using over the GPT-4. Because it only brings a small amount of enhancement, but the price increases a lot.

1 Like

that was also my takeaway from the aider benchmarking post:

Claude 3 Opus and Sonnet are both slower and more expensive than OpenAI’s models. You can get almost the same coding skill faster and cheaper with OpenAI’s models.

but i’m still super excited to play with claude 3.

what i really want is an autogen-style team with both gpt-4 and claude-3 on the team trying in parallel, and then a “manager” that decides whose output is better (not sure which model is the better manager).

2 Likes

Opus so far is great. Less hallucination.

Looking forward to the long context mode! Hope it lands soon!

1 Like

are you saying this is why there is no ‘c# LLM’ ? the language doesnt matter?

Interesting to see, but I think these benchmarks are far from the full picture

I see your point here. Very valid… but 10 a day and then incrementally slower for every other request is a bit of a stinger for paying customers in my opinion. It’s not going to go far at all.