The best model on the market isn’t winning the market. Anthropic’s top-tier AI is struggling to pull in users while cheaper tools pick up the slack, according to Marcus on AI, which flagged the story as fresh trouble for the frontier labs. The framing is blunt: more bad news for the companies betting everything on building the smartest model in the room.
What stands out here is the gap between quality and adoption. For two years the whole industry ran on one assumption. Build the best model, and users follow. That assumption is cracking.
What’s actually happening
Buyers are voting with their wallets, and they’re not paying a premium for the last few points on a benchmark. A model that’s 90% as good for a fraction of the price wins most real-world jobs. Summarizing email, drafting copy, answering support tickets, writing code snippets. None of that needs the frontier. It needs “good enough” and cheap.
Gary Marcus, who runs Marcus on AI, has been beating this drum for years. He’s the field’s most persistent skeptic, and he’s been early on calling out the limits of pure scaling and the shaky economics underneath the hype. Worth noting his track record cuts both ways: he’s been right that scaling alone wouldn’t deliver AGI, and critics say he’s too quick to write off genuine progress. On the business math, though, the signal is getting hard to ignore.
Why it matters now
The frontier labs spend billions training each new model. That only pays off if customers pay up for the best. If the market treats intelligence as a commodity, the whole cost structure stops making sense.
Three forces are squeezing at once:
- Price collapse. Token costs have fallen off a cliff. What cost dollars last year costs cents now.
- Open weights. Free and near-free models keep closing the gap on the paid leaders.
- Good-enough demand. Most business tasks don’t reward extra IQ. They reward low cost and low latency.
Put those together and you get a brutal question for Anthropic, OpenAI, and Google. Who pays a premium when the discount option does the job?
The Future Cast: where this goes by 2028
Expect the value to move up and down the stack, away from the raw model. Here’s the likely shape of the next two to three years.
- Models become interchangeable. Companies route each task to the cheapest model that clears the bar. Frontier models get reserved for the hard 10%.
- The margin moves to the product. Winners won’t sell intelligence. They’ll sell workflows, memory, integrations, and trust. The wrapper eats the model.
- Consolidation and pain. Labs that can’t fund the next training run on subscription revenue alone will lean harder on enterprise deals, raise fresh capital, or get absorbed.
What to do about it
If you’re building on AI, this shift is good news for your costs. Act on it.
- Don’t marry one model. Build a routing layer so you can swap providers as prices and quality change. Lock-in is the expensive mistake.
- Test the cheap tier first. Run your real tasks on smaller and open models before defaulting to the flagship. You’ll be surprised how often it holds.
- Compete on the product, not the model. Your moat is the workflow around the AI, not the AI itself. Own the data, the interface, and the customer relationship.
The race for the smartest model isn’t over. But the race to make money from it just got a lot harder, and the labs that assumed brains alone would sell are learning that the buyer doesn’t always agree. More detail on the numbers behind the story is available at the original source.