Good Enough Is Eating the Frontier

The best model rarely wins the workload, and Anthropic’s flagship is now the clearest evidence we have.

The Financial Times reported that Anthropic’s top model is struggling to attract users while cheaper tools thrive. The company’s overall growth remains strong. Adoption of the premium tier does not match it. Gary Marcus framed this as more bad news for frontier labs heading toward public markets, and the market-structure read is fair. But there is a more interesting story underneath, and it is architectural.

Capability is not the binding constraint

Most people still reason about model choice the way they reason about buying a laptop: pick the fastest one you can afford, use it for everything. That mental model breaks the moment you build an agent.

An agent is not one inference. It is a loop. Plan, call a tool, read the result, revise, call again, verify, retry. A single user request can fan out into dozens of model calls, and each of those calls carries the accumulated context of everything that came before. Cost does not scale with the number of tasks you run. It scales with the number of steps inside each task, multiplied by a context window that keeps growing as the loop continues.

That changes the economics completely. A premium model that is meaningfully better per call can still be the wrong choice for a system that makes four hundred calls, because most of those calls are not hard. Deciding which of six tools to invoke is not hard. Extracting a field from a JSON blob is not hard. Summarizing a file you just read is not hard. Paying frontier prices for those steps is like hiring a principal engineer to run your CI pipeline.

Where the marginal intelligence actually pays

The teams I see running agents in production have converged on a shape that nobody advertises: a small number of expensive calls at the decision points, and a large number of cheap calls everywhere else. The expensive model does planning, ambiguity resolution, and final verification. Cheaper models do the grinding.

This is routing, and it is now the central design problem in agent systems. It means the frontier model is not competing for the workload. It is competing for a thin slice of the workload, the slice where marginal intelligence changes the outcome. That slice is real and valuable. It is also much smaller than the total volume of tokens an agent burns.

Seen that way, weak adoption of a flagship model is not a verdict on the model. It is a verdict on how much of a typical pipeline genuinely needs it.

The evaluation gap nobody prices in

There is a second reason cheaper tools thrive, and it is uncomfortable for anyone who builds models for a living. Most organizations cannot measure the difference.

To justify a premium tier, you need an evaluation use that captures the failure modes the cheaper model produces and the expensive one avoids. Building that use is real work: task suites that reflect your actual domain, graders that do not just check string equality, enough volume to separate signal from variance. Without it, the procurement conversation collapses into a price comparison, because price is the only number both sides can see.

So the frontier model loses on a technicality. Not because it fails to deliver better reasoning, but because the buyer has no instrument that registers the improvement. Labs have been optimizing capability curves while the customer-side measurement stack lagged badly behind. The gap between those two is where cheap models live comfortably.

What this means for the labs

There is an internal signal worth reading against this. Reporting on Anthropic indicates that by May 2026, more than 80% of the code merged into the company’s own codebase was authored by Claude. That is a workload where the frontier model clearly earns its cost, because the users are engineers who can evaluate the output instantly and the task is exactly the kind of long-horizon reasoning that rewards a better model.

The lesson is not that the flagship is unwanted. It is that its value is concentrated in domains with tight feedback loops and expert users, and diffuse everywhere else. Frontier labs have been selling a general-purpose premium product into a market that increasingly buys specialized capacity.

For the IPO question, that matters more than any single quarter of adoption numbers. A business built on the assumption that the best model captures the most tokens is priced differently from one that captures the most valuable tokens while conceding volume to cheaper tiers. Both can be good businesses. They are not the same business, and the second one requires a story about routing, orchestration, and evaluation that the labs have barely started telling.

The uncomfortable version of this finding is that intelligence has become a commodity input with a demand curve, and demand curves respond to price. Advanced capability does not sell itself. It has to be cheaper than the alternative at the step where it matters, or measurable enough that someone will pay the premium anyway. Right now, for most of the work agents do, it is neither.

🕒 Published: August 24, 2026