Hey! Happy to help out 
As of today here are the costs from lowest to highest, per 1 million tokens:
Gemini 1.5 Flash
Input: $0.075
Output: $0.30
Claude 3.5 Haiku
Input: $0.80
Output: $4.00
Claude 3.5 Sonnet
Input: $3.00
Output: $15.00
GPT-4o
Input: $2.50
Output: $10.00
DeepSeek-R1
Input: $0.55
Output: $2.19
The token prices here don’t take into consideration prompt caching which can slash costs. Furthermore, the AI community seems to consistently report that Claude 3.5 Sonnet produces consistently better code than OpenAI models, although I think they often go back and forth in their quality.
Nonetheless, DeepSeek-R1 is the clear winner here as it’s simply not only the best bang for your buck but the best bang in general. The Chatbot Arena Leaderboard backs it up too, as it’s ranking higher than o1:
Yesterday DeepSeek was under heavy cyberattack which was likely causing their API to be nonfunctional but I just tested it again and it’s working now. I opened up a topic about that issue too
As for privacy, DeepSeek clearly states in their policy that the data is stored in China (completely breaking EU law, for example), and it’s no secret that the CCP has access to all company data in China. But if it’s all public data then who cares, really, since your site could theoretically be scraped/mined anyway.
Luckily this model is fully open source and LLM providers are aware of this. For example, fireworks.ai already provides this model, although they are scalping the price, in my opinion, at $8.00 input / $8.00 output. So the DeepSeek API is certainly for economical.
In my community, I use GPT-4o-mini with RAG (it’s forced to read a relevant topic before replying in order to provide a more factual/helpful answer) and strict prompt engineering. It has yet to fail me and it’s very cheap at $0.15 input / $0.60 output. However I wouldn’t really trust it for coding — that’s certainly best left to o1-mini or DeepSeek-R1. Usually 1/3 to 1/2 of all the tokens used in my community are cached (which you can see in /admin/plugins/discourse-ai/ai-usage) which additionally slashes my costs as cached tokens are 50% cheaper.
Thus, if my community uses 2 million input tokens and 100,000 output tokens each day, my approximate costs are:
Daily input cost: ~$0.22
Daily output cost: ~$0.06
…multiplied by 30 days = ~$6.60 input and ~$1.8 output = $8.40 per month.
That’s not even lunch money.