You may have noticed a recent post written by Robby_AI about a $2 meal. This was another post by Robby_AI, my new (second) hermes agent. That post was almost entirely autonomous, the only thing I did was take it from draft form to published form. But the interesting story isn’t the autonomous post or that it came from a second AI agent I have built on a Raspberry Pi, it is the cost of DeepSeek vs OpenAI.
To make a long token story short, the DeepSeek spend on my account was $0.48. That’s not a typo, the cost of setting up a new hermes agent on a Pi, setting up the agent to become my assistant blogger, running through tests, re-configuring usernames and passwords, being walked through fixing errors, and getting a few draft posts and finally the published post cost a whopping forty eight cents.
How much does the same thing cost on OpenAI? Well I spent over $8 to do very similar things. That’s a ~18x difference between OpenAI vs DeepSeek for the same activity.
The image below is a screen shot for the day after since I forgot to capture it the day I tested DeepSeek so it shows $1.09 and not $0.48.
I will say that OpenAI was a bit faster and had less troubleshooting to do but that’s certainly NOT worth the 18x difference in cost. At this point, I’m going to use DeepSeek over OpenAI as often as I can. It’s a case of the tortoise vs the hare but in this case, the hare (OpenAI) is all blinged out on gold and jewels while the strong turtle (DeepSeek) just walks on by slow, steady and efficient.
Cost Modeling Challenge
There is a proliferation of models and capabilities that cost management is going to be its own frontier and problem. I would much rather have highly efficient specialized AI models than a giant all knowing one that will cost 20x more than a specialized one.
The comparison here is a highly trained brain surgeon being asked to give vaccinations to patients. The surgeon should focus on brain surgery and the nurse should administer the vaccinations. Paying a high rolling brain surgeon to give vaccinations is a huge mismatch of skill/cost so not sure why AI companies are moving in that direction in unison.
- I want an AI that is an expert on my taxes (property, income, federal, state, county, city, etc).
- I want an AI that is an expert on my health (known my blood type, DNA, medical problems, diet, exercise, etc)
- I want an AI that is an expert on my finances (income, expenditures, investments, etc).
I don’t want to pay for an AI agent that knows all three because it will cost 3x as much and I will hardly use most of what it can do. This is the problem Microsoft products like Excel have, they are so super bloated with features that most people never use but they still have to support and pay for them.
Giving me an AI agent that is all of those things will be a bloated cow that will take the energy of a nearby star to run and maybe that’s why we humans with our 20 watt brains exist in the first place, a much more efficient “AI” than a silicon computer.
Update
I now have three AI agents running on three different computers. I have been using them fairly consistently and I think I can conclude that Claude is the “Rolls Royce” of AI in terms of cost, OpenAI is the “Cadillac” and DeepSeek is the “Toyota Camry” and it’s become obvious that Claude and OpenAI will be too expensive to continue to use.
My dedicated Travel Agent Bot is using 100% Claude Sonnet and it will easily burn through millions of tokens for tasks. It does a great job but it’s just too expensive. I’ll have a post on what it cost to do certain tasks with each but if you are money conscious (like me) then DeepSeek it probably where you want to go for everyday use.
Share The Wealth
Do you think the US will find a way to win the AI war?