The Price of Not Choosing

For three years, most organisations never chose an AI model. The subscription chose for them. That era is ending, and the adjustment it demands is not technical.

Brightbeam published a sharp piece this week on AI model routers, and it deserves engagement rather than a nod. Their argument goes like this. Routers, the software layer that decides which model answers which query, were until recently a tool for obsessives. Now that flat-rate subscriptions are tightening and pay-per-token pricing is becoming a lived reality, routing will go mainstream. And if the layer that directs AI workloads becomes as important as the layer that directed search traffic, routers might become the Google of the agentic era, capturing value by deciding where the work goes.

Half of that argument is right, and the half that is right matters a great deal. The other half is where I would push back, and pushing on it leads somewhere more interesting. It leads, as these things tend to, back to the organisation rather than the technology.

The satisficers were never lazy

Start with what Brightbeam gets right, including their best observation, which is that the prevalence of satisficers hid where market power might lie. For three years, most organisations using AI were what economists call satisficers. They picked a good-enough model, usually the most famous one, and sent everything to it. The blog post and the board paper, the regex fix and the regulatory summary, all of it went to the same frontier model at the same flat monthly price.

This was not laziness. It was rational behaviour under a specific pricing regime. When the marginal query is free, optimising which model receives it is pure cost with no benefit. The subscription made model selection a non-decision, so nobody built the muscle for making it.

That pricing regime is now visibly unwinding. Anthropic moved its newest model to metered usage credits this month. Google has shifted Gemini from fixed prompt caps to compute-based limits that drain faster on complex work. OpenAI restructured its tiers in April and pushed heavy Codex usage toward credits. Perplexity quietly metered a plan it once described as unlimited. The pattern across every major provider is the same. The flat fee increasingly buys the floor, not the ceiling.

The reason is arithmetic, not strategy. SemiAnalysis ran the numbers in June and found that a $200 monthly subscription, used to its theoretical maximum on long-horizon agentic work, would cost as much as $14,000 at standard API rates. By their estimates, some providers slip into negative gross margin on top-tier plans at utilisation rates in the single digits. Flat-rate pricing was a customer acquisition subsidy, and the bill for heavy use is being repriced to reflect what the work actually costs.

So the diagnosis holds. The marginal query now has a visible price, and once something has a visible price, someone in the organisation has to decide whether it is worth paying. That is new. That is the real story.

Where I would push

The prescription is where I think the argument gets ahead of its evidence, in three places.

First, the Google analogy. Google captured the value of directing traffic because it owned demand. Users started there, and it auctioned their attention to the highest bidder. Routers own neither the user relationship nor the workload. They sit between a customer who chose them and model providers who can disintermediate them, and they charge accordingly. OpenRouter, the category leader, was valued at $1.3 billion in May and now routes roughly 25 trillion tokens a week, which sounds like market power until you notice the business model. It is a take rate of about five per cent on inference spend that passes through, roughly $50 million in annualised revenue. That is a genuinely good business. It is a toll on flow, not a chokepoint on demand. And the hyperscalers already bundle the same function into cloud contracts, where it costs nothing extra. AWS Bedrock, Azure Foundry and Vertex are all routers in effect. A margin-compression layer is not a market-power layer, and the two should not be confused just because both sit in the middle.

Second, the claim that second-tier Western labs stand to gain if Beijing restricts its models. The Beijing part is real. Reuters reported this week that China's Ministry of Commerce has met Alibaba, ByteDance and Z.ai about restricting overseas access to their most advanced models, closed and open-weight alike. But it is worth being precise about what has actually happened. Meetings have been held, options sketched, nothing decided, and the restrictions under discussion may apply only to future models rather than anything already released. It is a live risk, not a done deal.

More importantly, the routing data tells you who actually wins if cheap Chinese tokens get scarcer, and it is not who the argument suggests. On OpenRouter, Chinese models went from roughly two per cent of routed traffic to around forty-five per cent in a single year, because they are cheap and good enough. The demand they serve is a price-driven commodity market. If that supply thins, the traffic flows to whoever wins the next rung of the price-performance curve, and the frontier labs' own small models sit on much of that rung already. Brightbeam names Mistral, Meta and xAI as the labs positioned for a resurgence, but that reads less like an analysis of routed demand and more like a list of Western labs that are not OpenAI or Anthropic. The traffic will follow the cost curve, not the flag.

Third, and most fundamentally, the tension the router thesis skates past. Routing works best where models are interchangeable, and the agentic workloads where the real money sits are precisely where they are not. The same routing data makes the point. Anthropic's models carry roughly twelve per cent of OpenRouter's token traffic but around forty-six per cent of its revenue. The market has already split in two. There is a commodity layer where price decides and switching is free, and a premium layer where reliability, tool use and lab-specific harnesses create stickiness that per-query arbitrage cannot touch. The more agentic the work, the less routable it is. Which means routers will capture the layer of AI spend that matters least, and the valuable layer will consolidate around whoever makes agents that actually finish the job.

The question nobody assigned

Strip out the overreach and something solid remains, and it is not really about routers at all.

The end of flat-rate AI forces a question that most organisations have never had to answer, because the pricing regime answered it for them. What level of intelligence does this task actually require? A Wall Street Journal analysis found that matching workloads to appropriate models can cut AI costs by as much as ninety-five per cent. That is not a software feature. That is a judgement about the work, and someone has to be authorised to make it.

Consider what that judgement involves in a regulated environment. A pharma company cannot route a pharmacovigilance summary to whichever model is cheapest this week. Model choice in a validated workflow is a controlled change, with documentation, risk assessment and sign-off. The router does not remove that governance burden. It multiplies it, because every routing rule is now a decision someone must own, and every model swap is a change someone must justify to an auditor. The technology to route queries in twenty-five milliseconds exists. The organisational machinery to decide what is allowed to be routed, by whom, under what quality threshold, mostly does not.

This is the deployment gap wearing a new costume. For three years the gap was between buying AI and using it. The next version is between using AI and governing its economics. Finance will want the ninety-five per cent saving. Quality will want the validated model. The business will want whatever finishes the task. Someone has to arbitrate, and in most organisations that role does not exist, because until this year the question did not exist either.

Brightbeam closes by asking whether the optimisers were on to something. They were, and that is the piece's real insight. But what the optimisers were on to was not a piece of software. It was a discipline, the habit of asking what each task is worth before spending on it. Software can route a query in milliseconds. It cannot tell you what the query was worth. That was always somebody's job. The flat rate just let everyone pretend otherwise.