How much intelligence do you really need?
Increases in model capabilities will be lost on most people now
Last week Anthropic released their latest model, Fable 5. I’ll save the drama around its release (and subsequent un-release) for the news bits section below. By all accounts, Fable 5 is a real leap in model capabilities. It seems especially good at autonomous workflows, with reports of it working unsupervised for 10+ hours on an open-ended goal and producing fantastic results. This has left many questioning how useful that level of intelligence will be for them. I think for most people the answer may be, “not much,” but that won’t stop model developers from making money. Cars provide the perfect analogy to help me explain.
A street legal golf cart costs about $15K. A typical commuter car like a Toyota Camry costs about $30K. A Ferrari 296 GTB, on the other hand, costs $350K.
The speed limit on almost all roads is higher than the top speed of a golf cart. The golf cart would often get you where you want to go, but most people will be willing to pay the extra money to get there faster. At the other extreme, the extra speed of the Ferrari is useless to most people. Those who drive a Ferrari are either doing so purely because they can, or because they actually need that speed for racing.
If every available car were slower than the speed limit, it would make sense to buy the fastest one you could afford. For a while, that was also my advice about AI models. Monthly access costs range from $0 per month to $200 per month. Up until recently, I had felt that you should pay as much as you feel like you could afford, because the added intelligence would be a real benefit to you, if you learn how to take advantage of it.
That’s starting to change. Back in February I started to play with GPT 5.4 and codex, OpenAI’s coding platform, shortly after it was first released. It did essentially everything I asked it to do at the time: I built some amazing educational apps with it (see my weekly App of the Week posts), and found many non-coding uses. I also played a bit with Anthropic’s Opus 4.6, and it seemed about the same in capability. Shortly after that, GPT 5.5 and Opus 4.7 were released. I definitely noticed a difference, but it was relatively marginal. The models were already reaching the point where the added intelligence just wasn’t that useful to me. Now we have Fable 5, and soon we’ll have GPT 5.6. I’m questioning at what point the decision for which model to use becomes more like the decision for which car to buy: rather than just paying as much as you can for the fastest possible car, you now have to do the research to balance cost, safety, comfort, size, etc. Speed is no longer a consideration, because every vehicle (other than a golf cart) can go as fast as the speed limit. Having a car that can go 300 mph doesn’t help you get to the grocery store faster.
I think it’s under-appreciated how important this is. For example, just a few weeks ago I wrote about how I worry that the variable cost of intelligence will increase inequities in the classroom. There’s no getting around the fact that wealthier students can access more competent models. I now believe that over time, this problem will largely disappear. Not because newer models won’t be more intelligent, but because at some point even the open source (i.e., mostly free) models will be competent enough for most routine school work. The more expensive levels of intelligence won’t give wealthy students much more of a competitive advantage over their classmates. If it hasn’t happened already, tomorrow’s models won’t be any more helpful for Chemistry homework than yesterday’s.
I don’t know if Fable 5 has crossed the line into models that are more intelligent than the average user will need. At the time of this writing it was taken off-line (see below), and I never got a chance to test it out. However, what is clear about Fable 5 is that the ratio of cost vs the marginal benefit of higher intelligence is getting more extreme. When it becomes available again it will only be temporarily accessible through a monthly plan. When that ends it’s cost will be proportional to usage, and then it may be significantly more expensive than the top subscription tier (model usage for monthly subscribers is highly subsidized). Some people will really benefit from the extra level of intelligence, and for them it will be well worth the price. However, for most the benefits may not outweigh the costs.
Fable 5 could be the first Formula 1 car of AI. Such cars cost upwards of $20 million, and only professional race car drivers benefit from their speed. The racing industry is profitable enough to justify these costs, but the average consumer doesn’t need those capabilities. When choosing an AI model, the question for most of us will no longer be “What is the smartest one I can afford?” It will be “What model is smart enough for the work I actually do?”
News Bits
It’s been one of the most dramatic weeks of all time in AI, and the news is all about Anthropic’s Fable. I’ll give here a brief timeline of recent events, with the caveat that there’s a lot we don’t know, and things are still unresolved.
For the last few months Anthropic has had a model called Mythos, which was reportedly far superior to any widely available model, and particularly good at finding (and possibly exploiting!?) security vulnerabilities. For that reason, it was only made available to security professionals (including within the Defense Department) through Anthropic’s Project Glasswing.
After months of developing safeguards for a wider release of Mythos, last Tuesday Anthropic released what they thought was a safe version called Fable 5. This included an automatic downgrade to Opus 4.8 when asked biological or cybersecurity questions.
On release, Fable also included a safeguard against people trying to copy it through a process called distillation. Unlike the downgrade to Opus 4.8, this safeguard silently kicked in without warning or notice, making the model untrustworthy for some legitimate AI research purposes. This resulted in some public anger from the AI community. On Friday morning Anthropic reversed their position, and subsequently when this safeguard was triggered the model would give notice.
Sometime Thursday night Amazon CEO Andy Jassey notified senior administration officials that researchers had “tricked” Fable 5 and Mythos 5 into revealing some cybersecurity vulnerabilities through a process called jailbreaking. Anthropic ultimately claimed that the information their model revealed was limited, and other publicly available models (e.g. GPT 5.5) could be similarly tricked to reveal the same information.
Friday night after 5pm the government tried to contact Anthropic CEO Dario Amodei. He could not be reached for 75 minutes. When he was contacted, he wanted more information about exactly what the jailbreak was and what information was revealed before he would do anything in response.
Later that evening the government issued an export control directive, telling Anthropic they must revoke access to these models for all foreign nationals both outside and inside the US, including those working at Anthropic. As Anthropic has no way to check citizenship, they said the only way they could comply was to revoke access for everyone, which is exactly what they did.
This is where things stand with Fable at the time of this writing. I’m sure there will be more to report on this next week!
David Bachman is a professor of Mathematics, Data Science, and Computer Science. He writes about AI and its real-world impacts. To learn more about his academic work, mathematical art, or AI speaking, consulting, and curriculum development, visit davidbachmandesign.com.