Gary Marcus put it bluntly when he flagged the Financial Times report: “Cheaper tools are thriving.” He framed it as more bad news for frontier AI companies, and specifically as a question about the IPOs those companies are marching toward. His post led with the FT’s finding that Anthropic’s best model is struggling to attract users.
My first reaction was not surprise. It was recognition. I have been building bots for a living, and I can tell you exactly where my API spend goes, and it is not the top of the menu.
What the Report Actually Says
The claim is narrow, so let’s keep it narrow. Anthropic’s flagship model has low adoption relative to budget options, even while the company’s overall growth stays strong. Cheaper alternatives are picking up the users. That’s the finding.
Notice what that is not. It is not a story about the flagship being bad. It is not a story about Anthropic shrinking. It is a story about a gap between capability rankings and purchasing behavior, which are two different scoreboards that people keep confusing for one.
Why This Matches What Builders Actually Do
Here is the shape of most bot architectures I have shipped or reviewed. There is a thin layer of genuinely hard reasoning, and then there is a mountain of boring work around it.
- Classifying an incoming message into one of six intents
- Extracting a date, an order number, and an email from messy text
- Rewriting a user query for a vector search
- Summarizing a retrieved document down to something that fits a prompt
- Deciding whether to escalate to a human
- Formatting a final answer into JSON that won’t break the frontend
None of that needs a frontier model. A cheap, fast model handles every item on that list, and it handles them at a cost per call that lets you actually run the thing in production without a finance conversation. When you build a pipeline with fifteen model calls per user turn, the price per token stops being a rounding error and starts being the product.
So the pattern that emerges naturally is routing. Cheap model by default, expensive model on escalation. That architecture is not a compromise. It is just correct engineering. And if a lot of builders converge on it independently, you would expect flagship adoption numbers to look exactly like what the FT described.
The Part That Should Worry Frontier Labs
The uncomfortable implication is about pricing power, not capability. If the top model is where your margins live, and the top model is where the users aren’t, the revenue mix gets awkward. Marcus is pointing at the IPO question for a reason. Growth that comes mostly from budget tiers is growth with a thinner floor under it.
There’s a second wrinkle. Capability improvements at the frontier tend to trickle down into the cheap tier fairly quickly. Last year’s expensive model is roughly this year’s affordable one. That dynamic is great for people like me and structurally rough for anyone trying to charge a premium for the ceiling.
What I’d Actually Change in Your Stack
If you take one thing from this news cycle, make it an audit rather than a vibe shift.
Instrument your model calls
Log which model handled each step, the latency, and the cost. Most teams I talk to cannot answer “what percentage of our spend goes to reasoning versus formatting” and that is a solvable problem with an afternoon of logging.
Downgrade one step and measure
Pick your least interesting model call, the one doing extraction or classification, and swap in the cheapest available option. Then run your eval set. If quality holds, keep it. If it drops, you learned where the hard part of your pipeline actually lives, which is more useful than a benchmark chart.
Keep the expensive path available
Do not rip out the frontier model. Route to it. Complex multi-step reasoning, ambiguous user intent, anything with a legal or financial consequence attached. The value of the top tier shows up in a small slice of traffic that matters a lot, and that’s a fine place for it to live.
Capability Is Not the Same as Fit
The broader lesson in this story is that being the most capable model does not automatically make you the most used one. Markets reward fit, and fit includes price, speed, and how well a model slots into a pipeline someone already built.
For those of us assembling bots out of small parts, that’s an easier world to work in. We get to treat models as components with different cost profiles instead of choosing a single champion and building around it. The frontier labs will keep pushing the ceiling, and I’ll keep using it sparingly, right where it earns its price.
đź•’ Published: