Deepseek v3 beats sonnet 3.5

[debian3]

It seems like there is a new model beating Sonnet 3.5

According to Aider new polyglot benchmark:

Any plan to add this model?

[debian3]

DeekSeek v3 is 53x cheaper to inference than Claude Sonnet

Now available on OpenRouter:

Deepseek V3 - API, Providers, Stats

Anyone have found more benchmark about it?

[SrGabrysh]

I’ve come across several benchmarks, but I find them overly optimistic. I also ran a few tests, and when generating text in French, I still encountered some parts written in Chinese, which is concerning.

Additionally, it’s impossible to register using a professional email address, just like on Kling.ai. I want to draw your attention to this issue because it raises some questions. Why are professional email addresses not accepted? With Kling.ai, I was never able to obtain an invoice, and I’m afraid the same might happen with DeepSeek. This could be a significant problem for businesses.

[woaidianqian]

In China, you need a real name to use it.

[SrGabrysh]

I have a real name, and my email address is based on my domain name, like any professional email address. It is unacceptable to be required to provide a personal email address to access the service. Even more concerning is the inability to obtain invoices, which is a major issue.

[debian3]

I have read several report on reddit that the model is quite good and extremely cheap. Anyone had a chance to try it?

[debian3]

According to LiveBench it score below Gemini 1206 and well below sonnet 3.5

[Kirai]

I tried it for a few hours with Roo Cline extension and I think the new whale is slightly behind Sonnet 3.5 (new), but at fraction of its cost. I would guess 90-95% Sonnet’s performance, but faster and if I remember correctly, only 7% of Sonnet’s cost (after the planned API price increase).

I believe it could be a nice model to have if it was “priced” in Cursor accordingly, e.g. one tenth of premium use (I think it should be possible, since new Haiku is one third?).

[debian3]

My guess at that price they could include it in the unlimited small model like 4o-mini

[raythurnvoid]

I tried it, it seems on par with sonnet in my opinion, but visibly faster.

Check this out to try yourself: Add DeepSeek Model · Issue #1509 · getcursor/cursor · GitHub

The problem is that you cannot use the composer with it, so even assuming it’s slightly better in some cases, it just cannot compete with Sonnet in the composer either normal or agent.

[qq1469617613]

Here’s the data I’ve seen

[alehano]

I tried DeepSeek 2.5 and it wasn’t good - slow and generated whole file instead of necessary parts.
But all changed when DeepSeek 3 released.
It’s fast and works great. At least as good as Sonnet 3.5 maybe even better.
I use deepseek-chat model it as my main model now.
But there is one issue here. Composer mode couldn’t work with this.

I really hope Cursor team will allow it in composer. And hope for community support to speed this process up

[truell20]

Testing it right now! Will definitely add it if it performs well on our qualitative tests / evals.

[Rain]

In fact, the reduction in costs and the increase in speed are good for both the cursor company and its users！

[Rain]

Of course, it would be better if the composer could open custom APIs to pro users, as this would give users more choices, and it would be worth more for pro users to purchase it.

[djbob2000]

Maybe it’s worth starting a beta branch/version so that you don’t have to wait for production, but can immediately have the opportunity to try new features?

[debian3]

I’m pretty sure the branch we are on 0.44.x is already beta.

[Rain]

Very good suggestion. Can a test version be opened for users to test?

[Rain]

I have been using the cline and curosr custom API with deepseek v3, and the results are astonishing!!!

[Rain]

Is there a new test result after a week, and will it be added?