Remember when everyone lined up to buy the phone with the best camera specs, then spent the next two years taking blurry photos of their lunch? The gap between what a product can do and what people actually use it for has always been wide. Right now, that gap is showing up in one of the most expensive corners of technology.
The Financial Times reported that Anthropic’s best AI model is struggling to attract users, even as cheaper tools thrive. Anthropic itself is growing strongly. But its flagship, the one built to be the smartest thing in the room, is seeing low adoption compared to budget options. Gary Marcus, who has been skeptical of frontier AI for years, picked up the story and pointed it at the obvious next question: what does this mean for the AI companies heading toward public offerings?
I want to look at this from a different angle, because I think most of the commentary is going to focus on stock prices and valuations. For those of us who just want to understand how AI agents fit into normal work, this story is actually useful. It tells us something about what people want from these tools.
Smarter is not the same as more useful
Here is the assumption that has driven a lot of AI spending: build the most capable model, and customers will come. More reasoning ability, better benchmark scores, longer context windows. Charge accordingly.
That assumption is being tested. If the strongest model is not pulling users while cheaper ones grow, the most likely explanation is boring and human. Most tasks people hand to AI are not that hard.
Think about what you actually ask an AI assistant to do in a given week:
- Summarize a long email thread you did not read
- Rewrite a paragraph so it sounds less stiff
- Pull the key dates out of a contract
- Turn messy notes into a tidy list
- Write a first draft of something you will heavily edit anyway
None of that requires a model operating at the edge of what is possible. A mid-tier model handles all of it. And when you are running thousands of these small tasks through an automated agent, the price difference per task stops being a rounding error and starts being your budget.
What this means if you build with agents
If you are using AI agents in your work, or thinking about it, the practical lesson is to stop shopping at the top of the menu by default.
A useful way to think about it: match the model to the task, not to the marketing. Sort your work into rough tiers.
Routine, high-volume, low-stakes
Classification, extraction, formatting, first drafts, tagging. Cheap models do this well. Run them here and check the output with simple rules rather than a more expensive model.
Judgment-heavy, low-volume, high-stakes
Complex reasoning, tricky code, analysis where a subtle mistake costs real money. This is where the expensive model earns its price. You will call it far less often than you think.
Plenty of teams building agents already do this. They route the easy steps to a cheap model and escalate to the strong one only when a task actually needs it. That pattern is quietly reshaping how AI gets bought, and it may explain part of what the FT is describing.
The uncomfortable part for the frontier labs
Training a top model costs an enormous amount. The business case for that spending assumes people will pay a premium for the result. If the premium tier struggles to attract users while cheaper alternatives thrive, that math gets harder to explain to investors, especially ones considering a public offering.
There is a second problem. Today’s frontier model tends to become next year’s cheap model. Capabilities that felt remarkable eighteen months ago now run at a fraction of the cost. If the top of the market keeps getting commoditized from below, being first to the frontier buys you a shorter window than the spending implies.
None of this means the strong models are pointless. Anthropic has said that more than 80% of the code merged into its own codebase was authored by Claude, up from earlier figures. That is a real signal about what these systems can do at the harder end of the work. But internal usefulness and market pull are not the same thing, and the FT report suggests the second one is proving trickier.
What I would take away from this
Advanced capability does not automatically translate into commercial success. That sounds obvious written down, yet a very large amount of money has been spent on the opposite belief.
For anyone reading this site, that is good news. You do not need the most powerful AI in the world to get value from AI. You need the model that fits the job, priced so you can use it as often as the job requires. The market appears to be figuring that out ahead of the headlines.
So the next time a launch event promises the smartest model ever built, the fair question is not whether it is impressive. It is whether you have any task that needs it.
🕒 Published: