Comment by baalimago
12 hours ago
New Deepseek models are like Christmas for me. Really big fan of low cost API models, noone does it better than DS. Until VRAM price is low enough to run models locally, this is the way to go.
The subsidized subscription model won't last, API pricing "feels" closer to a true sustainable business model.
Indeed. My fellow software engineers keep complaining about using up all their Claude tokens within an hour... Whilst I'll be rocking DS flash for the entire day. Sure it gets a few things wrong here and there, but that's when you pull out the Claude models or whatever for those tricky tasks.
Same. And I have come to use OMP (oh my pi) agent /advisor mode to put a 2nd model on the case (also mid-size one), reading everything. It can not block anything or change anything - just inserts comments in the text stream with 1 turn delay. Good portion of the time it's quiet. I'd say 1/2 of the time it's got something to say. About 2/3-rd of that the 'advice' is insubstantial or about something not-quite wrong. The good thing is the main model is confident - checks and then it stands its ground. Have not noticed it turning a right into a wrong b/c of advisor false alarm. And in 1/3-rd of the advice, it's a genuine defect teh advisor noticed, the main model works out a fix. This is my approximate feeling just observing the process, have not got collected the data. Afaik only OMP has advisor mode. Agent pi has plugin pi-omplike-advisor. For agent Hermes I had them code me an /advisor plugin (for now -0.1 old v0.18.x; yet to upgrade it to latest).
The problem is picking between models. I do not want to spend my time switching models and trying to decipher which model should be used for what. Maybe that's just a me problem that I need to figure out.
DeepSeek v4 is honestly good enough that I'm fine throwing it at everything in my hobby projects. I guess now I'll be switching from V4 Pro-Preview to V4 Flash. My only real complaint is that they can't do images, which limits their ability to autonomously debug some kinds of issues
Of course you can get more bang for your buck by being more deliberate. But that's equally true with US frontier models. You can optimize your work by choosing between Opus, Fable, Sonnet, Sol, Luna and Terra for each task. Some people seem to prefer to let Opus code and Sol review, for example. And then there is the whole debate whether $current_version is actually better (Some people stay on Claude 4.8 because they dislike how 5.0 is sometimes doing stupid things, just as many opted out of dynamic reasoning when they still could)
1 reply →
what plan are your 'fellow software engineers' using? I have a hard time even using up the Fable part of my allowance in a week of coding.
I'm not sure to be fair, but they do have constant "token anxiety", which I simply don't have anymore since using v4 flash.
4 replies →
I'm on the Premium business plan and I can easily use up my entire week's allocation in two days of coding. I only use Claude for planning, too, the rest is done by Deepseek and GPT.
Claude is very spendy.
2 replies →
usually I am - Codex, make pi with deepseek to do something.
I would bet that Deepseek API pricing is still more cost effective per token than the subscriptions. With the increase in quality Deepseek Flash just got (in my personal testing so far, it seems to have improved a lot at following instructions, and has become more proactive), there really isn’t anything that can match it in terms of cost effectiveness.
The issue for me is the data privacy if you're using their hosted prices, because you cannot opt out of data collection (and I'm not sure how you'd legally follow up on it if they did offer it, but didn't actually follow through). That means all of my code is being retained for training.
I've used it in some open source code though, and loved how fast it was.
My mind is changing on how valuable my code actually is though... it's the complete picture, how it's put together, the design, the UI, the attention to detail that's the real value.
I'm not a heavy user of agentic coding, but still use them quite a bit for some automation here and there. I've been going around shopping all the ~$10 subscriptions and I finally settled on openrouter + ds4 pro. The more intensive days cost me $1 and I set a $2 weekly limit which I've never hit the past 3 weeks, to me it's way cheaper than most subscriptions and I don't have to worry about maximizing my weekly quota/resets.
I have one of my projects on DeepSeek and it blows my mind how I can use tens of millions of tokens for pennies.
I don't find it suitable for everything, but there's some tasks it crushes for what feels like almost free.
FWIW, looking at pricing myself, it needs to 5x the tokens to get the same results as GPT 5.6 Luna at a much lower TPS. (doesn't obviate the fact that the pressure causes incumbents to push and push to optimize, thank you Deepseek!)
[flagged]