How to price AI-driven features (LLMs are expensive to run), any ideas?
Hello Indie Hackers!
I've just rolled out a new feature on https://www.smart-investor.co/ that enables investors to instantly analyze a company's earnings call with a single click.
Link to the demo: https://x.com/dayem_d/status/1718251601327009839?s=20
The Problem We're Solving 🤔
Investing requires diligence, but let's face it, not every individual investor has the luxury to sift through lengthy earnings calls like a Wall Street analyst. This AI-powered feature aims to bridge this gap by offering:
📝 Quick Summaries: Don't have time to listen to the entire call? Get the gist quickly.
📊 Key Metrics: Identify business KPIs that matter the most.
😃 Sentiment Analysis: Gauge the overall tone—positive, negative, or neutral.
⚠️ Risk Assessment: Understand any potential red flags you should be aware of.
The goal is to let investors have a quick glance at key insights from earnings calls before deciding to delve deep into them. This way, you can efficiently allocate your research time.
Pricing Challenge 🏷️
AI-driven features like this come at a cost. While the expense is manageable for now, as usage grows, so will costs. I've implemented caching to avoid re-running the LLM model for the same document, mitigating costs to some extent.
I'm contemplating how should I price this feature/product. I think a fixed monthly subscription fee should scale well as more people use the service. I'd really like to avoid a credit-based pricing system where customers have a limited number of AI queries.
Seeking Your Opinions 🙏
What are your thoughts on the utility and design of this new feature?
Any suggestions for pricing models that both cover costs and offer value?
Any additional suggestions and opinions?
Hello Dayem, how about pricing based on the value your product is offering rather than the services that you're using to run it, I can't see a reason why someone who isn't heavy cash-wad would want to shift through lots of financial data, so how about charging just a little bit below your traditional competitors, it will still even be more than the cost to run, this way you can then even afford a free-tier so prospects can use your product first.
Hi Dayem,
We are running into a similar issue and we are currently considering changing the pricing to tiered subscription models to a usage based pricing.
as for keeping the model usage cost low, we started off using GPT4 but in the past few weeks, we have been using our finetuned GPT3.5 turbo model. Our results are the same in terms of output but the cost and the speed of generation have significantly reduced.
I faced the same dilemma a few months ago before launching my side project (otheryou). In my situation, the costs are not "manageable for now". I utilize several resource-intensive models in the backend to process my videos (RVC, Roop, GPF Gan, etc.). It takes almost 30 seconds to process just 1 second of video on a T4 GPU. Thus, processing a 2-minute video costs me around $1.
Without external funding, it's unrealistic to offer my users a free trial or similar incentives.
As a result, I opted for a token-based system: $10/month equates to 480 tokens (~8 minutes of videos). This decision is risky for the project. The absence of a free plan might deter many potential users, hindering the product's chances of going viral, especially since it's geared towards entertainment too. Only time will tell if this was the right move and if my product finds its market fit.
At the end, if a product can't recover its costs, there's little justification for its existence.
A credit or usage-based pricing model is the only way to go IMHO. Your revenue really needs to be directly aligned to your costs so things are predictable and manageable, otherwise you're just assuming all the risk and paying Open AI regardless of whether your model works or not.
A monthly subscription would work well, but you'll have to define a limit to avoid mis-use. Otherwise, the credits system is the way to go. Other than that, you can also consider using a vector database to lower the number of back/forth queries, but how to use this depends on your own implementation.