Your model has a six-week shelf life

Look at what shipped this month. In a single two-week window the frontier moved with Claude Mythos 5, GPT-5.6, and Gemini 3.2 — and then a wall of Chinese models landed almost on top of each other: Qwen 3.7, DeepSeek V4.1, GLM-6, and several more. This isn't a busy month. It's the cadence now: a new frontier model roughly every six weeks, from multiple labs, indefinitely.

That cadence changes what a model is to your product. It's not a foundation you pour once. It's a consumable with a shelf life.

"We use the best model" is not a strategy

If the thing that makes your product good is that it runs on whichever model topped the benchmarks this month, you have a problem with a calendar attached. The model you picked is, on average, six weeks from being matched and a few months from being cheap. Your competitor doesn't need to be smarter than you — they just need to swap to next month's model before you do.

A capability that everyone gets in the next release isn't a moat. It's a countdown. The benchmark lead you're proud of is the most perishable asset you own.

Build on the swap, not the snapshot

The teams that handle this well stopped treating "which model" as an architecture decision and started treating it as a routine, reversible one. The model lives behind a boundary; switching is a config change and an eval run, not a project.

That's the same lesson Apple just shipped to a billion phones by making the assistant's model a setting. When the part most likely to change is isolated behind a clean interface, a six-week cadence is an opportunity — you upgrade the moment the numbers justify it. When it's welded into your system, the same cadence is a treadmill of migrations you'll never finish.

What actually compounds

If the model depreciates every six weeks, put your durable effort into the parts that don't:

• Your data and context. The retrieval, the grounding, the proprietary information the model reasons over. That's yours, and it gets better while models churn.
• Your evals. A test suite on your tasks is what lets you adopt a new model in a day instead of fearing it for a month. It outlives every model it judges.
• Your orchestration and product. How work is decomposed, how tools are wired, how the experience feels. That's where a real edge lives, because it doesn't ship in anyone's next release.
• The swap itself. A clean adapter so a new model is a line of config. Cheap to build once, priceless every six weeks.

None of those show up on a model leaderboard. All of them are still yours after the leaderboard reshuffles.

The bottom line

The release cadence isn't slowing down, and no single model is going to stay on top long enough to be a foundation.

Treat the model as a consumable with a six-week shelf life: build the swap, and invest your durable effort in data, evals, and product — the things that compound while models churn. The frontier will move again next month. Build so that's good news.