The Model Layer Is No Longer Where the Value Accrues

For most of the last few years the rational move was to pick the strongest model and route everything to it. The gap between the best model and the second-best was wide enough that paying the premium was obviously correct. That gap is closing.

The Latent Space piece states the mechanism plainly: intense competition among frontier labs combined with the ever-increasing power of open-weight models like Kimi K3 and Qwen 3.8 Max has made routing a key part of deployment. When a free-to-run open model is within a few points of a metered frontier model on the tasks you actually do, sending every request to the expensive option is no longer discipline. It is waste.

The size story reinforces this. Z.ai's Jie Tang argues that parameter count alone is no longer a useful shorthand, and that there are now two 2-3 trillion parameter open-weight models chasing a frontier estimated at 3-7 trillion, separated by only a couple of points on the AA index. A couple of points. That is the entire premise of routing: if the quality delta is small and the cost delta is large, the correct model is different for different tasks.

So routing is not a clever optimization bolted onto a settled market. It is the direct consequence of the market fragmenting. When there was one obvious best model, you did not need a dispatcher. Now you do.

Aggregation Theory says platforms win by aggregating demand and then commoditizing supply. Whoever owns the user relationship wins, and the suppliers behind the interface get squeezed into interchangeable inputs. Applied to AI, the models are the supply. The router is the aggregator.

The Stratechery framing of the OpenRouter deal makes this explicit: it reads the acquisition as an implicit bet on a future market of models and the chance at Aggregation. Read that carefully. The bet is not that any single model wins. The bet is that no single model wins, and that in a fragmented market the layer sitting between demand and dozens of interchangeable suppliers captures the value.

Stripe is a natural buyer for exactly this reason. It already owns a demand-aggregation position in payments, and metered model access is, structurally, another metered-usage business. The same source notes the deal involves flipping the business model, which is the tell: the interesting question is never the routing algorithm, it is who gets to sit in the toll booth and on what terms.

The frameworks here point the same direction. Commoditize Your Complement: if you own the router, you want the models below you to be cheap, plentiful, and substitutable, which is precisely what a strong open-weight field delivers. Every capable open model that ships makes the model layer more of a commodity and the routing layer more of a business. The labs are, in effect, subsidizing the aggregator above them.

Wardley Mapping tracks components as they evolve from genesis to commodity: novel, then custom-built, then productized, then utility. The useful discipline is asking where a given component sits today and what that implies for where the money moves next.

Eighteen months ago the frontier model was a product you selected, integrated, and differentiated on. It sat in the product stage, maybe still edging out of custom. The routing trend is the visible sign that it is sliding rightward toward utility. When Jie Tang says the meaningful questions are now how much data you have, where you spend your compute, and who runs the model under what conditions, he is describing a component being unbundled into operational variables. That is what commoditization looks like from the inside.

The supply-side flood confirms it. The Sequence describes a single week as another four-model week, with DeepSeek V4-Pro reaching general availability, Z.ai shipping GLM-5.3, and NVIDIA releasing two models at once, dryly noting that this is four opportunities to lose an afternoon comparing decimals. When capable models ship weekly and differ by decimals, no single one of them can anchor a durable product position.

Here is the Wardley move that matters for readers: value does not disappear when a component commoditizes. It relocates to the adjacent component that is still evolving. The adjacent component here is the layer that orchestrates across the commodity below it. Routing. That relocation is the whole story.

This publication's Harness Hypothesis holds that the value in AI is not in the model but in the harness that connects the model to the world. Model routing is the harness growing a decision-making organ. It is the harness deciding, per task, which brain to borrow.

You can see the harness taking on this weight in the tooling. Claude Code's recent release notes include a fix for prompt caching for sessions using an LLM gateway or custom base URL, which is a small, boring line that quietly assumes you are not talking to one vendor's endpoint. You are talking through a gateway. The presence of gateway plumbing in a first-party tool is a signal that multi-model, routed operation is now the default assumption, not an edge case.

The agent frameworks tell the same story from another angle. Pydantic AI's changelog includes a fix to not send Anthropic thinking blocks with an empty signature, the kind of provider-specific reconciliation work that only exists because a single harness is expected to speak to many models with different quirks. The harness is absorbing the complexity of a fragmented supply market so the user does not have to.

The user-facing consequence is the part worth internalizing. When you choose an agent now, the model behind it is increasingly a variable the harness sets, not a constant you own. Your experience is shaped less by any one model's ceiling and more by how well your agent's dispatcher matches tasks to models. Two people running the same agent on the same frontier model can get materially different results if the routing behind them differs.

Routing would matter on quality alone. It matters far more because the cost side of AI is under real pressure that has nothing to do with token pricing negotiations.

Latent Space's memory coverage is blunt: prices are up 500% in twelve months, with 128GB DDR5 kits ten times the lowest price ever seen, a squeeze they nickname the RAMpocalypse. Memory pressure flows straight through to inference economics. When the hardware that serves models gets dramatically more expensive, the premium for sending everything to the largest frontier model stops being a rounding error and starts being a line item you notice.

That is the environment routing was built for. If a task can be handled by a smaller open-weight model at a fraction of the serving cost, routing it there is not penny-pinching. Across an agent that runs thousands of steps a day, it is the difference between a viable deployment and an unaffordable one. This is why Glean's enterprise business leans on routing and why the trend is described as equally hot in enterprises, where usage volume makes the arithmetic unavoidable.

The practical framing for a power user: your agent's monthly cost is now largely a function of its routing policy, not just its headline model. An agent that intelligently downshifts routine work to cheap models and reserves the expensive frontier for genuinely hard tasks can cost a fraction of one that routes everything to the top. The dispatcher is the budget.

There is a quieter consequence that the market narrative skips. Every model a router can reach is a place your data can go. Multi-model routing multiplies the number of destinations for your prompts, your context, and whatever the agent has access to.

The Trust Boundary Model says to identify every point where data crosses from one trust level to another, because those crossings are where you inspect and enforce. A router is, by definition, a stack of trust boundaries: each candidate model may be a different vendor, a different jurisdiction, a different set of retention terms. Sending a task to an open-weight model self-hosted inside your walls is a different trust proposition than sending it to a third-party frontier API, even when the routing decision that chose between them is invisible to the user.

The sandboxing conversation running through the ecosystem is the flip side of the same coin. Researchers are actively probing how to run untrusted code with no network access and constrained filesystem access, and there is a live argument that modern sandbox primitives lower deployment cost and provide good security boundaries for safely extending software. That work matters precisely because routing and extensibility are expanding the number of trust crossings faster than governance is keeping up.

The uncomfortable synthesis: as routing becomes the norm, the question of which model saw your data becomes as important as which model produced your answer, and it is far less visible. The dispatcher that saves you money is also the dispatcher deciding where your context travels.

Put the pieces together and the definition of choosing an agent has quietly changed. You are no longer choosing a model with a personality. You are choosing a routing policy with defaults.

The forces are all pointing the same way. Open-weight models are within a couple of points of the frontier. New capable models ship in four-model weeks. Serving costs are inflated by a memory squeeze. And the smart money, in the form of Stripe's reported $7B, is buying the aggregation layer rather than any single model. When those conditions hold, routing is not a feature. It is the product.

For readers, three questions now matter more than 'which frontier model does this agent use.' First: can I see and control the routing policy, or is it a black box that quietly decides for me? Second: what is the cost profile of that policy over a month of real usage, not a demo? Third: which models can my data reach, and do I trust all of them, not just the flagship?

The consensus keeps scoring the race by model benchmarks because that is the number that is easy to publish. The real contest has moved one layer up, to the dispatcher that decides which of those benchmarked models ever touches your task. That layer is now the infrastructure. Watch who owns it.