The AI market has a lemons problem — that's why nobody picks just one model

I keep noticing the same pattern in teams doing serious AI work: nobody uses just one model anymore. A cheap, fast model handles the high-volume, low-stakes calls — classification, extraction, first-pass drafts. A frontier model gets reserved for the handful of calls where being wrong is expensive. A specialized model handles code. None of this is arbitrary. It’s a deliberate routing decision, revisited every few months as the options shift underneath it.

Call it model optionality: instead of standardizing on one vendor’s stack, power users treat model choice the way a portfolio manager treats asset allocation — diversify, because betting everything on one position is the actual risk.

Why the market makes that the rational move

The obvious answer is cost and capability both move too fast to commit to. Price-per-token keeps falling as more frontier-capable models arrive from more vendors, while raw capability keeps climbing — the two curves that would normally justify a long-term platform bet instead make one look reckless. A model that was the clear best choice two quarters ago can be the expensive, outdated option today, and the vendor with the best price this quarter might not hold that position by the next.

Enterprises that architected around a single provider’s API quirks and prompt patterns are now paying a switching tax to get out. The ones that built for portability from the start — thin abstraction layers, workload-level benchmarking instead of vendor loyalty — are the ones actually capturing the falling prices instead of getting stuck holding a shrinking capability lead.

The deeper reason: nobody can fully verify what they’re buying

But cost and capability churn alone don’t fully explain the hedging instinct. The real driver sits underneath both: buyers genuinely can’t verify a model’s quality before committing to it at scale. Benchmarks are gameable. “Production-ready” means something different depending on which vendor is claiming it. A model can look excellent on a demo and on a leaderboard and still fail in ways that only show up on your specific workload, months in.

That’s a textbook information asymmetry problem — the seller knows more about true quality than the buyer does, and without a reliable way to close that gap, the market starts to resemble George Akerlof’s “market for lemons”: buyers, unable to tell a genuinely strong model from one dressed up well, start pricing and choosing defensively. Nobody wants to be the firm that bet its whole pipeline on this quarter’s lemon.

Model optionality is the practical hedge against exactly that risk. Running the same task across two or three models and comparing outputs on your own data is a cheap, ongoing quality check that no vendor’s marketing page can substitute for — and it means a bad model degrades one lane of the pipeline instead of the whole thing.

What that means if you’re the one deciding

If you’re choosing how to deploy AI in a firm right now, the strategic question isn’t “which model is best” — in a market this volatile, that answer expires before you finish implementing it. The better question is whether your architecture can absorb being wrong about that answer: can you swap a model out in a week, not a quarter, when a better or cheaper option shows up, or when the one you’re using turns out to be worse than it looked.

This is the second entry in The Long View*‘s Business & Economics section — ongoing essays on how AI is reshaping competitive strategy, as I keep digging into where the real leverage is.*