Look at the state of play. SWE-bench Verified is topped by Claude Mythos at 93.9%, gated to a handful of partners, so no one outside that circle can build on it. Terminal-Bench Hard is topped by GPT-5.5, ahead of everything Anthropic currently fields. The best SWE-bench number a normal team can actually call is Opus 4.8 at 88.6%, now the Claude Code default, with a fast tier that just dropped to a third of its former price. Three benchmarks, three winners, and not one of them is the answer to which model the product should use. Capability stopped being a scalar this week. It became a shape.
The default existed because it was rational. For most of the last two years one model was clearly ahead, switching cost was high, and every provider's surface was different enough that supporting two was real work. Pinning the best one you could afford was the correct call. It was a bet that one model was best at everything, and for a while the bet paid because one model nearly was. That is the condition that ended. The frontier is now a set of specialists wearing a generalist's name.
So the decision is not which model. The decision is routing. Which model takes the terminal and CLI work, where a competitor is measurably ahead. Which model takes the long-horizon reasoning. Which model takes the cheap high-volume path that runs on every request and cannot cost what the frontier costs. Which model takes the latency-sensitive surface, now that a fast tier exists at a price a product can actually pay. Cost, latency, and capability are three different axes, and no single model sits at the top of all three. A default collapses that space to a point and throws away everything the other models were better at.
A default is a bet that one model is best at everything. No model is best at everything anymore.
This is why routing is a design decision and not a config value. The model a request lands on determines what the product can do, what it costs to do it, and how it feels while doing it. Those are the things the product team is supposed to own. Hand routing to a default in a settings file and you have let an infrastructure convenience decide the product's behavior. The eval set is what should decide routing. The set already encodes what right means for each kind of task. Run it across the candidate models and the routing table falls out of the results. The team that writes the evals is the team that should own the routing, because the evals are where the product's standards already live.
The failure mode is quiet, which is what makes it expensive. A pinned default never throws an error. It just returns a slightly worse answer on the tasks its model is worst at, every time, and the aggregate metric stays high enough that no one looks. The product ships the wrong model for a third of its traffic and reports an acceptable average. The gap does not show up until a competitor who routed by task is visibly better at the exact thing your default model is weakest at, and by then the gap is a quarter old.
The pattern is the one this journal keeps finding. RAG became architecture once the storage decision stopped being load-bearing. The design system became substrate once the components stopped being a craft project. The agent became a contract once the loop became a library call. Model choice is the next thing to move, from a default you set once to a routing decision you draw on purpose. The studios that compose a product out of several models, each handed the work it is built for, will draw a different product than the studios still pinned to one.
Field note prompted by the early-June 2026 cluster: the SWE-bench Verified and Terminal-Bench Hard leaderboards splitting between vendors, and Claude Opus 4.8 shipping as the Claude Code default with a fast tier at a third of its former price. The principle outlasts any single leaderboard.