The AI boom has been built on a basic assumption: Bigger models are more powerful, and the most powerful models win. Now, the industry is about to learn what happens if that assumption starts to break.  

Mounting costs have already pressured users to give smaller and cheaper models a second look. This cost-conscious model-shopping is new and it’s unclear how it will affect the industry, but the impact is likely to be significant. 

One prediction, laid out best by Coinbase co-founder Brian Armstrong, is that it will result in the vast majority of tasks shifting to cheaper models. 

“[D]emand for intelligence is near infinite, but 80% of workloads will be running on 99% cheaper models within 12-18 months,” Armstrong wrote on X. “20% of workloads will still run on latest gen models where IQ maxing is important.” 

It’s hard to overstate what a significant shift it will be for the AI industry if Armstrong’s prediction comes true.  

Before now, most AI companies have competed on quality, which has meant defaulting to the most advanced available model. If those same jobs can be handled by cheaper models without affecting quality, it would mean a massive shift in the economics of AI. And critically, much of the savings would be coming out of the pockets of the big labs, dealing a financial blow to OpenAI and Anthropic just as they’re heading for their IPOs. 

It’s a potentially seismic change in the industry, resting on one basic question: Are companies ready to switch to smaller models? 

Initial tests suggest that, when the system is arranged right, cheaper models could sub in without any sacrifice in quality. In a recent test by the legal AI tool Harvey, the company was able to reduce inference costs by 3x without reducing quality. The test, performed in partnership with the inference platform Fireworks AI, combined Claude Opus and Fireworks’ GLM 5.1, and shifted to Opus for the most intensive tasks. The result was a significantly lower load in terms of server time and overall cost. 

“Quality comes first, and in legal it always will,” Harvey co-founder Gabe Pereyra told TechCrunch, referring to the AI legal services his startup provides. “However, the definition of quality is evolving from simply using the most powerful model for everything, to using the best model that gets the right answer most efficiently.”

This trend is often framed in terms of major labs versus Chinese models or open-weight ones, but that misses the bigger point. The real divide isn’t between proprietary and open models; it’s between large models and small ones. You can save money by switching from GPT-5.5 to DeepSeek’s V4 Flash, but switching to GPT-5.4-mini works just as well.  

There’s an active price war going on between in-house inference from the big labs and independently served open-weight models. For the bigger question of small versus large, it doesn’t really matter which kind of small model wins out.  

All of this might seem obvious — of course you shouldn’t use more compute than necessary — but it runs counter to the scaling-first approach that has dominated the industry until now. Inspired by the bitter lesson, labs have leaned hard into training the most compute-intensive models possible, pushing the frontier of what AI models can do. With prices heavily subsidized by investors, clients had no reason to choose anything but the most advanced option.

With token prices rising and subsidies slowing down, users are facing cost pressure for the first time. We don’t know whether the new cost pressure will actually drive enterprise users to smaller models. They could just as easily economize by making fewer calls, using less context, or simply giving up on the least promising deployments. 

But if it turns out that most deployments can be run just as well on a smaller model, it could put a serious damper on the growing demand for inference — and raise new questions about how to justify the cost of training a frontier model. 

Anthropic has released Claude Fable 5, the first publicly available version of its closely watched Mythos model. What can Fable actually do? All kinds of things, it turns out.

Ethan Mollick, a notable AI researcher and University of Pennsylvania scholar, has been playing around with the model and seems to be having a lot of fun.

In his testing, Fable consistently “outperformed basically every other public model I have used by a considerable margin,” Mollick wrote Tuesday on his Substack. He added that it was “capable across many problems and produced some startling results — it would work up to a dozen hours executing on multi-page specifications.” 

Perhaps most strikingly, Mollick used Fable to create a variety of video games — all of which were generated via “one initial prompt” in Claude Code, the researcher says.

Among these, Snake is exactly what it sounds like. You’re a Pac-Man-like snake and you roam around eating apples. The snake never stops moving, and if you run off the screen, you die. It’s very 1980s arcade but, like many of those old games, it’s weirdly addicting. I played it longer than I’d like to admit before remembering I am a gainfully employed writer and not, in fact, a serpent who likes fruit.

Then there was Strata, where you’re roaming around in a seemingly endless network of subterranean tunnels and the goal is just to light as many lanterns as possible. The graphics look like a degraded version of Myst — they aren’t great — but the fact that the game exists at all, generated from a single prompt, is impressive.

Mollick even managed to create Duino, a game based on the Duino Elegies, the celebrated cycle of poems by poet Rainer Maria Rilke. I like the animation here best — the player is a lone figure in a nocturnal landscape — although there isn’t much to the gameplay other than walking around while Rilke passages materialize on the screen.

Aside from the variety of instant games Mollick produced, he also used Fable to create an isochronic map — a visualization showing how long it takes to travel between any two locations. The accuracy and detail is arresting.

The implications are pretty clear. Software projects that once required entire teams — games, mapping tools, highly complex specifications — are now being spun up from a single prompt. It’s reason for vibe coders of the world to rejoice. As for founders and operators watching AI capability curves, it’s a useful data point about how quickly the floor is rising.

Two years and a $250 million lawsuit later, Apple’s AI Siri revamp is on its way to your phones and laptops and even your mixed reality headset, if you happen to be one of like three people who actually uses the Apple Vision Pro. Apple revealed a slew of new information at Monday’s WWDC keynote about these long-awaited, AI-powered updates that can take advantage of the fact that our hardware is supposedly “built for Apple Intelligence.”

To be honest, it’s hard for AI to impress me enough that I’ll use it in my day-to-day life. I still don’t trust LLMs to provide consistently accurate information, I find it ethically untenable (and uncool) to use AI to help me write, and I don’t feel the insatiable urge to know what I would look like as a Studio Ghibli character. But every once in a while, the promise of AI tempts me.

That’s how I felt watching Apple’s Siri AI demos, which depict a world where your phone comes with an always-on, constantly-working assistant who knows everything about you and can help you keep track of all of the conversations happening on like 12 different apps on your phone at any given moment.

To paraphrase Katy Perry, it feels so wrong (what are the privacy implications?), but it also feels so right (I am so overwhelmed by my phone and am begging for help parsing it all).

I want Siri to be my own personal Emily from “The Devil Wears Prada” — a “second brain” that anticipates my needs before I even know what they are. I want Siri to read my texts and automatically make an event when a friend and I decide we’re going to meet up for dinner on Thursday. I want Siri to remind me when I’m walking past CVS that I have a prescription ready for pickup. If I forget to reply to an important work email, I want Siri to remind me that I didn’t write back yet.

Siri AI won’t be able to do all of that out of the box, but it’s moving in the right direction. In one example at WWDC, Justin Titi, an Apple senior director working on AI engineering, asks the smart assistant to remind him of the dessert that his daughter mentioned recently. Siri searches across Titi’s phone to find a text from about a month ago, when his daughter mentioned that she wanted to make coconut cookies. It’s simple, but asking Siri to find that message saves time, rather than scrolling up through an entire month of conversation looking for that one specific text.

The new-and-improved Siri is designed to use “personal context,” which refers to any information you put into Apple-native apps, like iMessage, Notes, Calendar, Mail, Photos, and more. Siri will also be aware of what’s on your screen, so for example, if you scroll past a picture of a nice park on Instagram, you can ask it to find out where that park is. (We still don’t know if Siri will be able to integrate into non-native Apple apps; it seems like it might be up to the developers to make that happen.)

There already are apps like Poppy and Poke that try to create this kind of mobile, agentic AI. But the paradox of these AI personal assistant tools is that you have to give up a lot of personal data and privacy to make them work correctly, which may just cause you more trouble (remember that time when a Meta researcher ran OpenClaw and accidentally deleted her entire inbox?).

I can’t say that I love giving any tech giant my personal data, but Apple at least seems to care more about security than the other FAANG (MANGOS?) companies. On-device AI will always be more secure and less energy intensive than cloud computing, since the data is processed directly on your phone. (This is how current Apple Intelligence features like email summaries and AI emojis are generated.) But for the more complex tasks that Siri will confront, Apple pioneered private cloud compute (PCC), a way for devices to parse complex data over the cloud without even exposing your data to Apple itself. (If it’s possible to hack PCC, it hasn’t happened yet, even though Apple offers a $1 million bug bounty.)

In a recent conversation with the writer Calvin Kasulke — who is so internet-brained that he wrote a novel that takes place exclusively on Slack — I confessed what feels like a taboo desire to outsource all of my “life admin” to an AI.

“When you talk about the nonsense of the tech detritus in your life… I think the question is, ‘Is all that you have necessary?’ If it is necessary, isn’t it worth cultivating the skill and spending the time to do it?” Calvin told me. “I don’t think that those are skills that one should allow to atrophy.”

He makes a good point: Maybe instead of asking Siri to remind me about the TV show that my friend told me I should watch, I could pay more attention when I’m talking to my friends. I don’t want to get into the habit of forgetting more consequential details from my conversations.

“I’m sorry, but all of the commercials that are like, ‘What if I had the computer buy my kid a birthday gift?’ I’m like, ‘What if you learned what your kid likes?’ … Like, I don’t know man, it sounds like [they] don’t want to do the fundamental act of being a person,” he said.

Maybe when I say I want Siri to be like Emily from “The Devil Wears Prada,” I should remember that Emily’s character is on the verge of a crash-out. I know I can’t psychologically impact Siri like Miranda Priestly damaged Emily, but will I become the kind of person who can’t function without the friendly robot voice in my phone? Do I want to be that person?

At least if I decide to opt out from all of this, Apple will make that possible. Unlike Google’s controversial Search overhaul, the new AI Siri can be toggled on and off, so you don’t have to use it. Until then, I’ll have to decide if it’s worth it to taste the forbidden fruit of Siri AI.