Comment by bakugo
6 days ago
Prices have been very obviously trending up, not down. Even open weights models are becoming more expensive with every release. Computer hardware is ballooning in price.
6 days ago
Prices have been very obviously trending up, not down. Even open weights models are becoming more expensive with every release. Computer hardware is ballooning in price.
Prices are going up for BETTER quality -> not for the SAME level of quality.
People are willing to pay more for BETTER quality.
You obviously haven't seen DeepSeek v4 Pro's pricing if you think pricing only goes up...
Maybe so, but that becomes irrelevant when you consider that the new, better quality instantly becomes the expected baseline. So the price of the "baseline" quality is going up regardless.
Let's look at GPU prices as an example. Around 12 years ago, I bought a GTX 970 for around $350. That was considered a very good GPU at the time. Today, the "equivalent" GPU model (RTX 5070) now costs almost double. Of course, the newer GPU is much more powerful (more than double, in fact), but all the things you'd use a GPU for have also advanced and now expect an entirely new level of performance as a baseline, such that the older GPU is fairly worthless today. So most people agree that GPUs in general have become more expensive.
Regarding DeepSeek's price: it's obviously subsidized, and unlikely to match the actual inference cost right now.
Just wait for the next model and the next model architecture. Just wait for it, bro.
Gemini 3.5 flash is 25% cheaper than 3.1 pro, and outperforms it on almost every benchmark, most by a pretty wide margin...
It's still 5x more expensive than 2.5 flash
Cool.
There has never yet been a new model which actually improved over the previous ones. They suck just as much, and in the same ways, as the models of 3 years ago.
Grab a 5090 and run Qwen 3.6 35b on it (6 parameter seems to work best for me).
Then buy $10 (or $2, if you're cheap, and they take PayPal) of DeepSeek credits.
Whilst you're at it spring for a Claude subscription too and GPT.
Switch models between Qwen, DeepSeek Flash, DeepSeek Pro, and you can meet 99% of your code generation needs.
Hop over to Opus 4.7 (or 4.8, but I haven't really used it yet) and GPT-5.5 when doing very complex architecture/design or troubleshooting something where DeepSeek Pro is getting stuck.
It is ridiculous how cheap this stuff is now. It's affordable at third world prices.
None of that is cheap.
> spring for a Claude subscription too and GPT.
You started with some random pricing then veered off into impractical hand waving. Far above third world prices...unless you count the USA as third world, I guess.
The extra subscriptions are optional. You can do nearly all of it with just a DeepSeek subscription and switch between Flash and Pro.
If you have the $$, do the extra stuff. People who like to play video games often have a very fancy graphics card that sits idle during their work day.