Three frontier models landed this month. I changed one line.

Look at the release calendar for this month. OpenAI is widely expected to ship GPT-5.6 — there's already a Codex log referencing it and Polymarket odds near 89% for a June release. Google confirmed Gemini 3.5 Pro is "coming next month" at I/O. Anthropic said Claude Mythos ships "in the coming weeks." That's three frontier models from three vendors in one month — and they're landing on top of Claude Opus 4.8 and Gemini 3.5 Flash, which both shipped in May.

For a lot of teams, a calendar like that is stressful. Every release reopens the same anxious meeting: is the new one better, should we switch, how much work is it, are we falling behind? For my own products, this month's three-model flood meant something much calmer. I read the announcements, ran my evals, and changed a config value. That's the whole story. And the gap between those two experiences is entirely an architecture decision you can make too.

The treadmill is optional

The stress isn't caused by the models. It's caused by how the code is wired to them. If a specific model's name, API shape, and quirks are threaded through your codebase — referenced from dozens of places, with prompts tuned to one vendor's exact behavior — then every new release really is a migration project. You have to go find all those references, change them, re-test, and pray. No wonder it feels like a treadmill: each model that ships is another lap.

But none of that is inherent. It's a choice to weld yourself to one provider, and it's an avoidable one. The whole point of the model being a commodity behind an API is that you can put a clean layer between your app and whichever model is behind it — and then swapping models stops being a project and becomes a setting.

Build the seam, collect the upgrades

The pattern is boring and it has a name now: a model gateway, or a provider- agnostic layer. Your code talks to one interface. Behind it, a thin adapter maps to whatever provider you point it at. Tools like OpenRouter and LiteLLM exist entirely to be that seam — one description puts it well, that OpenRouter abstracts model providers the way Stripe abstracted payment processors, so your application never hard-codes a single endpoint. You don't even need a library; thirty lines of your own indirection does it. The mechanism matters less than the property it buys: changing models is a one-line parameter update, not a refactor.

Once that seam exists, the whole calendar inverts. GPT-5.6 ships? Point the config at it, run your eval suite, keep it if the numbers improve, revert if they don't — ten minutes, not a sprint. Gemini 3.5 Pro turns out cheaper for your easy steps? Route the cheap work to it and keep the expensive model for the hard parts. Three releases in a month becomes three chances to get better or cheaper for almost no effort, instead of three fire drills. The churn that punishes the welded-in team rewards you.

Why this only gets more true

The cadence is not slowing down. We just watched the time between frontier releases compress to weeks, and with three labs leapfrogging each other every month, one analysis put it bluntly: single-provider lock-in has never been more expensive than it is right now. The faster models improve, the more value sits in being able to adopt the improvement instantly — and the more it costs to be the team that can't.

This is the part of "which model wins is a design smell" that's easy to miss. The point was never that the models don't differ — they do, and this month's three will each be better at something. The point is that the durable engineering decision isn't picking the winner. It's building so that picking is cheap, reversible, and ongoing — so you're never married to last month's choice.

The takeaway

The real moat was never a model. It's the seam. The seam is the thing a competitor can't download, the thing that turns a relentless release calendar from a source of anxiety into a stream of free upgrades you collect on autopilot. Everyone gets the same models; the team that built the clean layer between their product and those models gets to use the best one available on any given week, for the cost of editing a line.

So when the next three drop — and at this rate it'll be next month — the question isn't "which one do we bet on." It's "how fast can we try all of them." If your honest answer is "minutes," you've already won the only race that compounds.