Marks the next era of AI

Claude Fable is good. Like really, really good. It's also a sinister sign of where it's all going. The Uber moment. Or rather the frog in a pot situation.

This exact moment, July 7th is the beginning of a new era. It's the third most important shift in AI after the 2017 Google Paper and 2023 Chat GPT release.

LLM power-users are now the proverbial frog in a pot, and nobody seems to be ready for it. If you're unaware of the shift, you may be in for a nasty surprise.

Let's talk about what's going on.

Boiling the frog

There's a (rather untrue) saying: "boiling the frog". What it means is that if you gradually increase discomfort, you can safely get that discomfort extremely high. The frog will not jump out of the water if it slowly reaches the boiling point. Its fate is sealed.

The thing is, that frogs actually do jump out when it's uncomfortable. AI users likely won't. They will be boiled and it has already started. The Fable pay-per-token switcharoo is the first step of what's about to happen.

Short story

I'll start off with the high level prediction right away, and then we'll cover what it's similar to and how to prepare.

On July 7th, Claude Fable, the most powerful Anthropic LLM will move to pay-per-token billing only. Meaning it won't be a part of a subscription anymore.

And it is rather expensive.

Fable 5 lists at $10 per million input tokens and $50 per million output tokens. That's double of Opus 4.8 which has $5/$25. Fable is officially the most expensive model on Anthropic's current price list.

What I believe it means is a slow phase-out of subsidized AI. After all every $100 AI subscription is realistically costing up to $4000 in tokens (in some estimates). There are no "optimistic" scenarios in which those AI firms are profitable on subscriptions. The most optimistic is just a smaller loss.

Years ago Uber offerred ultra-cheap, subsidized cab rides. They simply paid the drivers out of their (investors) pocket. This is just a market dominance strategy. After that come price hikes, feature downgrades and nerfing.

And we're in one of those shifts right now. My quick take is this:

Fable will for a while be an expensive "extra" payment if you want to do something more complex and powerful. People will use Sonnet 5 and Opus 4.8 (or 5) for most regular tasks.

But just like with all other models, with time you'll start noticing your Opus is getting dumber. To avoid frustration you'll be switching and paying for Fable more and more often.

That is the boiling the frog part. The "subscription" models will get less and less powerful, while Fable will likely be the only model growing in output quality.

Give it a few months and doing anything serious on a $100 or $200 a month subscription will be pretty pointless.

That is such a bad take! Anthropic can't do that because they'll alienate their users!

Yes they can and they likely will. Those "paying users" are all costing them money. There is no profit coming from them. It's much better to have less users but have most of them on the powerful pay-per-token model.

Claude Fable was already "nerfed" since its July 1st reintroduction compared to the version from June. This is normal practice.

This wraps up the subscription subsidies

If (or rather when) this becomes reality, it will be the "Uber is now expensive" moment that will catch most people off guard.

This is what I meant when I said we have weeks left to vibe code. Any usage that doesn't have measurable ROI will stop making sense. You either use AI, pay for it and make money, or you don't use it at all.

Nobody buys a full tank of gas to just keep the car running at idle until it runs out. You need to get the most out of your investment.

When it was cheap, like $100 per month, it made sense to try out new things. To play around. To build another revolutionary to-do app.

Even Sam Altman had said, intelligence will eventually be metered, like electricity or gas.

The ROI problem

The problem with ROI from AI is that most businesses are still unable toshow there is one. There's some growth in some specific niches, but most way outspend any benefits.

Prominent studies from MIT and McKinsey reveal that between 80% and 95% of companies realize little to no bottom-line financial impact from their AI use

And with raising prices and dropping model quality that divide will become way more significant.

AI for Design ROI

We're still figuring this out in our respective industries.

I found a couple of use-cases that AI is perfect for in my design practice. They're not generative AI though, mostly data processing and parsing. Diagramming edge-cases.

I talked about parts of my process before and I plan to cover them in more detail soon.

There's an easy math here. Right now I can use it to expand multiple edge-cases and have more data covered. This saves me about 25% of the time, or rather I can deliver 25% more to the client in the same time, for the same money.

But once those costs grow by a certain margin, it won't be viable to use AI for that anymore. I'll either stop offering that extra work, or start doing it manually and raise our prices.

It's inevitable.

Local AI

Local AI can be an answer for specific use-cases. But it's not yet there for the kind of dev work people do using Opus, and especially Fable.

I use local models to parse data and output summaries based on very precise instructions, and that's as far as current local models work on 64 GB of ram. It works great for us at Pageformance, but only because it happens async. I am able to just let it run for 18 hours, with multiple local model checking each other's outputs.

It's a pretty sophisticated setup.

The problem with local is that if they get good enough, they may just as well destroy the Frontier AI companies. Dario Amodei has already complained about the dangers of Open Source (and open-weights) models.

Some predict Fable level models should be able to run on "consumer hardware" in about two years. That is a nice dream to have.

So we either get AI as a commodity to run locally for free, or powerful AI will be controlled and limited to the rich and powerful.

While option one feels amazing, have you ever seen the powerful not having their way in the end? Exactly.

I'm not optimistic about frontier-level local models. They'll likely ban the open source companies somehow if they get too powerful.

OpenAI to the rescue?

Just as Fable becomes paid-per-use, OpenAI plans to release their SOL model. Apparently it's their Fable level reasoning machine.

One thing to note is that they finally took a page out of Anthropic playbook regarding the naming. People don't associate with random letters and numbers. They much prefer things like Opus, Fable, Mythos.

Or Sol, Terra, Luna. That progression is also clear and more human. I like this direction from the design perspective too.

At first OpenAI will include Sol in their subscription. Obviously that is just a temporary and it will be heavily capped use anyway.

They have the extra cash to subsidize this a little longer. Anthropic won't be happy and they may lose some clients, at least temporarily.

In the end both Fable/Mythos and Sol will be pay per token, with subscription models decreasing in quality.

What it all means

I see all these pieces connect. The Uber price-hike moment is upon us. Starts Today. Fable will get better, Opus and Sonnet will get worse. At some point nearly unusable.

Sol will get better. Terra and Luna will get worse. At some point almost unusable.

This is how they boil the frog. You'll be switching to the more expensive paid model more and more without noticing it too much. And then in the end you either switch to it completely, or realize at these prices there is simply no ROI from using ai in your work.

This is why that's such a transformative moment. The economic part of the AI bubble is nearly ready to burst. We will have AI in the future, don't worry.

But it won't be on the terms we are using it now.

I for one used that one week of Fable to catch up on all my products security and stability. Did full run throughs for both web and mobile projects.

What was a commodity for the masses is slowly becoming a luxury for the select few.

What's your plan?