Hacker News Reader

Comment by VonLuderitz | Hacker News Reader

This month I changed my Claude Max Subscription to DeepSeek API. V4.1 Flash is fast and perfect for everything and refuses to almost nothing.

Hope DeepSeek starts month plans soon.

View on HN

> Hope DeepSeek starts month plans soon.

Do they really need since they are already cheap, make apparently profits with interference and API costs follow the usage? Monthly plan doesn't have much benefits for them, if most would be power users anyway? Only reason is to get more users on the platform, but people already get there, anyway.

Start monthly plans... so we can pay _more_? Lol. I've moved >80% of my daily coding usage to deepseek-flash, and at their current token pricing it's effectively free. Any monthly plan would almost certainly increase the cost.

They use your data for training though. Mostly I'm fine with that, but you do need to be mindful of what traffic you send their way.

There are cheaper ZDR providers on OpenRouter.

Quality is not as good in a lot of cases and the 10% worse cache negates any savings. Native Deepseek has the best caching of any inference provider.

OpenAI/Anthropic does not train on your data? I'm sure they do

You could consider going with OpenCode Go for 10$. The value you get is claimed to be 80$ in DeepSeek v4.1 Flash API value. I use that and so far it is good for me. Some claim they run a different quant but I so far don't see that.

Dax has said time and again they don't quantize the models they offer. Assuming they trust their upstream providers (I don't think believe they run their own inference) and we trust them, what you get with OpenCode Go is exactly what you'd get from the DeepSeek API.

I am personally a very happy OpenCode Go user. $10 month has never felt more useful.

Eh there's a lot of ways to 降智, quants are just the most visible. I noticed a fair bit of variance with opencode go and went back to deepseek direct.

It's helpful that DeepSeek is so cheap that it's actually reasonable to pay standard API rates.

This is no longer true(?), there's zero difference between what is served by deepseek and opencode go. Infact I am pretty sure it's served by deepseek as the tok/s graphs closely align with eachother.

I recently got Claude Max 5x, after using ChatGPT Pro 20x for a while, and I find the Claude Max value really good. For $90/m, the weekly usage is great AND you get $100 of API credits per month too.

Why the downvotes, lol?

I do believe Claude Max 5x gives best quality/value at the moment out of all plans/providers.

Probably because it’s a discussion about DeepSeek and you didn’t include that in your comparison so it came off a bit big-lab-fan

I didn't include it because the use-case is different.

I can't trust models like V4.1 Flash to do stuff for me without supervision. Yes, they are good for coding and some API-only processing tasks, but not as a main agentic model like Opus 5.5 or GPT-6 Astra are, which can actually do things that function properly without much fidgeting or intervention.

Also, a model like Haiku is better, faster and cheaper:

https://aibenchy.com/compare/anthropic-claude-haiku-5-5-high...

*on one set of private benchmarks

Artificial analysis has them benching similarly on intelligence, but with deepseek 4.1-flash having a decisive victory in time-to-first-output-token and end-to-end-response-time [1].

Combined these make the interactive experience of using deepseek 4.1-flash significantly more enjoyable than haiku IMO.

[1] https://artificialanalysis.ai/models/deepseek-v4-1-flash-101...

Anyone who compares DeepSeek to Claude can't be taken seriously. I'm rooting for open-source models of course but DeepSeek is so much worse than Claude, it's not comparable.

Out of curiosity, which harness do you use? Hermes, ..?

>refuses to almost nothing

What do you do to make it not refuse cybersec related activities? Sometimes it refuses to even scan my own codebase for vulnerabilities which is bizarre.

You mean Claude? I have had zero problems with DeepSeek

Reply on news.ycombinator.com