I spend over $100 per month on AI tools. It's over $1500 per year, not including one-off credits I'm buying to test new apps or temporarily ramp up API usage. I might be unique in my spending habits, but I doubt it. Because what I've learned in testing AI tools over the past year is that many of the best models and features are stuck behind a paywall. (If you want the full list of tools I pay for, it's available at the end.)
And as AI continues to advance rapidly, more and more companies are putting their models behind a subscription. This "subscriptionization" of AI is changing how we access and use these powerful technologies, presenting opportunities and challenges for users, businesses, and society as a whole.
The paid LLM landscape
To take a closer look at the state of AI subscriptions, I want to examine the three leading chatbots: ChatGPT, Gemini, and Claude. Each of their respective companies (OpenAI, DeepMind, and Anthropic) offers a paid version of the chatbot, offering access to flagship models and extra features. While each company markets itself as having a unique, state-of-the-art model, you'll generally find that they have more similarities than differences.
GPT-4: ChatGPT Plus provides access to GPT-4, which offers more accurate responses and fewer errors compared to GPT-3.5. Currently GPT-4 has a limit of 40 messages every 3 hours.
DALL-E 3: The subscription includes access to DALL-E 3 for image generation.
Extra tools: With GPT-4 also comes real-time internet browsing via Bing, data analysis/code execution, and vision and voice capabilities.
Custom GPTs: Subscribers can access and create custom GPTs from the GPT Store, allowing for tailored experiences.
Price: $20/month.
Gemini 1.0 Ultra: Like the others, the most advanced Gemini 1.0 model (Gemini 1.0 Ultra) is only available to consumers via a Gemini Advanced subscription. That said, there is also a Gemini 1.5 model, available in the "Pro" size, currently available to developers via a waitlist.
Integration with Google Apps: Gemini Advanced integrates with your Gmail, Docs, and Drive (with more apps coming soon), letting you chat with the model without leaving your current tab.
Extra tools: It's worth noting that the chatbot can also use existing Google services like Flights, Maps, and Hotels, can generate images*, and can work with voice and vision capabilities - even though they're not exclusive to the paid version of Gemini.
Price: $19.99/month (with a two-month free trial).
Claude 3 Opus: The Pro tier comes with access to Claude's latest and most advanced model, Claude 3 Opus.
Priority usage: Subscribers get not only 5x more usage than the free tier, but also have their chats prioritized during high-traffic periods.
Price: $20/month
Of these three, Claude brings the fewest features to the table for a paid subscription. However, Anthropic has said they intend to add more tools to Claude in the coming months.
But these are just three out of thousands of AI products that charge a monthly subscription. And though these LLMs are general-purpose, most AI models and AI-integrated apps are tailored for specific tasks: writing, coding, music, meetings, voiceovers, etc. As additional competitors enter the space, we'll only get more choices - and more subscription fatigue.
How I learned to stop worrying and love the free market
Of course, the rise of AI subscriptions isn't happening in a vacuum. Across almost every part of work and life, we're now inundated with a growing number of monthly subscription options, leading to "subscription fatigue."
I'm reminded of the pattern we saw play out with streaming TV companies. Although Netflix briefly reigned as the dominant streaming service, it now exists alongside Hulu, Disney+, HBO Max, Amazon Prime Video, Peacock, Paramount Plus, and more. What used to be a simple decision of whether or not to pay for streaming has now become a complex calculus of choosing which service to pay for during any given month.
The AI industry may follow a similar trajectory - at least for consumers. For software developers and for the open-source community, things are a little more of a mixed bag. On the one hand, companies like Mistral have released some of the highest-ranked open-source models freely and are looking to charge users for hosting and deployment services. On the other, Stability AI recently introduced a paid membership - meaning the only developers who can use their models commercially are the ones paying a monthly subscription.
To be clear, I don't begrudge any of these companies from charging money. Training a foundation model is incredibly expensive, and unlike most other software, AI consumes meaningful amounts of electricity and dollars to run at scale. If these companies want to exist beyond their latest VC funding round, they’ll have to figure out a way to recoup their costs1. Otherwise, our only recourse comes from deep-pocketed companies like Meta, which can spend billions on AI and release the results for free, thanks to their enormously profitable advertising business.
That said, the alternative would arguably be worse: a single AI company with a monopoly on GPT-4 level AI, with little to no open-source competition. Plus, the AI arms race does mean these tools are getting into the hands of more people, faster, for cheaper (though there are some key safety concerns too). With that in mind, I can make peace with the fact that we're in the middle of an awkward technology shift and assume that as things mature, making decisions will get easier.
I also hold out hope that market forces can find other ways of getting around our subscription fatigue. That could mean creating "bundles" with access to multiple AI tools for a single price or industry-wise standards (or regulations) that create easy interoperability to switch between LLMs. But until that materializes, it's ultimately up to you and me as consumers to determine what's worth paying for.
Choosing your AIs
While there's no right or wrong decision in choosing the best tools for you, there are a few things worth considering. And if you're a fellow AI enthusiast with plenty of disposable income, feel free to ignore my suggestions and pay for everything you feel like experimenting with!
It's probably worth paying for a leading LLM, and the leading ones are all good. The gap in capability between GPT-3.5 and GPT-4 is very noticeable if you're an LLM power user. While it might make sense to use the free version of ChatGPT/Claude/Gemini to get your feet wet, if you use it regularly or for work-related tasks, consider paying for the upgrade. And don't stress too much about which one to use - right now, they're all pretty close in performance. ChatGPT has been the pack leader for over a year, but Gemini has an advantage when integrating with real-time search and private Gmail/Docs.
Most other (text-based) AI services use ChatGPT/Claude/Gemini behind the scenes. While there are some exceptions, a lot of AI-enabled products “just” feed custom prompts and data to ChatGPT et al. That's not necessarily a bad thing - but make sure you’re fine with paying a “convenience fee” to have those features in the app, versus copying/pasting your own prompts. For example, I use ChatGPT to help me brainstorm, and have intentionally avoided paying for "AI brainstorming" products.
Decide if you actually need AI for images/audio/video. LLMs are general-purpose tools. But what about the plethora of specialized AIs for image, music, or video generation? It's up to you to figure out whether those tools make sense for you - I'm optimistic about AI, but I would also say that the vast majority of people don't need a subscription to Midjourney or Eleven Labs, even though they're both great products. If having an AI tool speeds up your work, or unlocks capabilities that are useful to you, consider experimenting. But this is going to be a very individual decision!
It’s been said before, but we’re in the middle of a Cambrian explosion of AI - and that’s a big part of why things feel a bit awkward. When cars first went mainstream in the US in the early 1900s, there were over 250 car brands. By 1929 there were only 44, and from the 1960s onwards, we mostly had the “Big 3”: Ford, GM, and Chrysler.
In the coming months and years, we will inevitably see some consolidation in the AI space. Everyone is rushing to slap an “AI” label on their product or company right now. But when the tide inevitably shifts, and the hype cools2, I suspect we’ll see a lot of smaller AI companies get acquired or shut down. The internet has created a lot winner-take-all markets, and I don’t see why AI would be any different.