Tokens Aren't Fungible

Open-source LLMs are roughly 90% cheaper than closed-source models at the same benchmarked intelligence level.

Yet they capture less than 30% of token share.

If intelligence were the only thing users cared about, the cheaper model should win. If price were the main thing users cared about, the cheaper model should win even faster.

It has not.

That is the puzzle. The LLM market looks like a commodity market from far away — tokens in, tokens out, benchmark scores, price sheets — but users are behaving as if each token carries hidden attributes. Reliability. Trust. Workflow fit. Safety posture. Privacy. Brand. Ecosystem integration. Latency. Support. The stuff that is hard to summarize in a leaderboard.

The puzzle: the cheap intelligence does not clear the market

The standard commodity intuition is simple: at the same quality, the cheaper supplier should win.

That intuition is powerful because it is usually right. If two cloud providers offer identical compute, the cheaper one should take share. If two exchanges offer identical liquidity and execution, price matters. If two models are equally intelligent and one costs a fraction of the other, why would users keep paying up?

The opening chart makes the puzzle visible: open models sit below closed models on price at similar measured intelligence.

But the quantity side refuses to follow the price chart. Open-source models remain below 30% of token share. OpenRouter even describes this as an equilibrium.

So the question is not whether open models are good. Many are. The question is why the market keeps paying for something beyond benchmarked intelligence.

Primer: six ideas the argument depends on

The central decomposition is:

Benchmarks, context length, reasoning flags, open-source status, and price live on the observed side. Workflow fit, trust, policy acceptance, support quality, internal switching cost, and brand risk often live on the latent side.

The naive model: intelligence is a commodity

The naive model says users buy intelligence. If two models have the same measured intelligence, they are close substitutes. Price should do most of the remaining work.

That model predicts three things:

1. open-source share should rise when open models get much cheaper at the same intelligence;
2. price should explain a large share of token demand;
3. benchmark-adjusted prices should converge over time.

The evidence points in the opposite direction.

The LLM market is not telling us intelligence does not matter. It is telling us intelligence alone is not enough to describe the product.

The failure mode: benchmarks measure what is measurable

Benchmark scores are useful because they are legible. That is also why they are dangerous.

A benchmark can tell us whether a model answers a certain class of questions well. It cannot fully tell us whether a customer trusts the provider with sensitive data, whether the model behaves predictably inside a brittle workflow, whether policy refusals are acceptable, whether uptime is good enough, whether procurement likes the vendor, or whether the support team will pick up the phone.

Those attributes are not soft. They are product quality.

This is why equally intelligent tokens can have different economic value. A token from a model that fits the workflow, passes internal risk review, and has predictable behavior is not equivalent to a cheaper token that creates operational uncertainty.

The better model: latent demand dominates the residual

Asset pricing has a useful analogy. In a demand-system view of markets, investors demand stocks for both observable characteristics and unobservable preferences. Koijen and Yogo call the unobserved component latent demand.

That language maps surprisingly well to LLMs.

Observable characteristics include:

• benchmark intelligence;
• price per input/output token;
• open vs closed source;
• context window;
• reasoning capability;
• model age;
• provider identity.

Latent demand includes:

• reliability and uptime;
• safety and policy posture;
• privacy and data retention;
• toolchain integration;
• support and SLA quality;
• brand risk;
• internal trust;
• switching costs;
• weird workload-specific fit.

The cited market evidence says intelligence alone explains only about 10% of pricing variation. Adding observable controls helps, but a large residual remains. The residual is not noise to ignore. It is the market telling us where the product actually lives.

Evidence check: the market keeps revealing hidden preference

The original essay walks through price dispersion, quantity concentration, regression results, and model stickiness. The exact numbers matter less than the pattern: measured variables explain some behavior, but not enough.

This is the part investors should sit with. A lower price can be real and still fail to move demand. A higher benchmark score can be real and still fail to win a workload. A model can be technically impressive and commercially second-choice.

The glass slipper: when fit beats frontier rank

OpenRouter’s “glass slipper” framing is a good way to understand the extensive margin.

A user does not necessarily search continuously for the globally optimal model. More often, they have a specific workload with awkward constraints. They try models until one fits. Once it fits, the model becomes embedded in prompts, evals, data pipelines, monitoring, cost expectations, latency budgets, product behavior, and internal trust.

At that point, switching is no longer a spreadsheet exercise.

The prize is not to be everyone’s second-best model. The prize is to become someone’s production default.

What this means for AI underwriting

If tokens are not fungible, then AI market underwriting should not stop at benchmark charts.

The diligence target shifts from “which model is smartest?” to “which provider owns hidden demand attributes that are durable?”

What to verify

• Reliability: Does the model behave predictably enough for production?
• Workflow fit: Is it embedded in a high-value, repeated job?
• Switching cost: What would break if the customer moved models?
• Trust: Does the buyer believe the provider is safe enough for sensitive use?
• Policy posture: Are refusals, moderation, and safety behavior acceptable for the workload?
• Support: Does the vendor help when the system fails?
• Brand risk: Would the buyer be comfortable defending the provider choice internally?

These are not easy to measure. That is the point. The unmeasured attributes may be where the margin lives.

A practical decision rule

The claim to watch

The market keeps saying the same thing in different ways: the benchmark is not the product, and the token is not the unit of differentiation.

The product is the full bundle that gets a user from uncertain task to trusted output. Some of that bundle is measured. Most of it is not.

That is why tokens are not fungible.