On June 2 at Build 2026, Microsoft shipped MAI-Code-1, its first in-house coding model, into Copilot and VS Code. Google routes its agents through the new Gemini Enterprise Agent Platform. OpenAI’s Codex is in the same lane. Three days later, Anthropic put Claude Opus 4.1 on a retirement clock: deprecated June 5, switched off August 5. Four vendors, comparable coding models, one week: the layer got more crowded and more disposable at once.
Why It Matters
That combination is the signal. When four vendors ship competing coding models in a single week, which model you pick stops being a differentiator. And Opus 4.1 was not pushed out by the market. It was retired on the vendor’s calendar, with 60 days notice, and requests to it fail after August 5. Opus 4 and Sonnet 4 go the same way on June 15. The model you standardized on is a leased component on someone else’s depreciation schedule. What you built around it is the part you keep. And the most durable piece of that layer is your review structure, the human-in-the-loop checks that catch when a new model quietly drifts from how the last one behaved.
The Decision
The question this week is not which coding model is best. It is where your AI investment sits: in the model a vendor can switch off on its own timeline, or in the layer that survives the swap: your context, workflows, and review steps. A reasonable CEO can standardize hard on one model, or stay deliberately swappable. Both are defensible. Confusing them is not.
What To Do This Week
- Ask which single model your core AI workflows are pinned to, then find that model’s published retirement policy. Every major vendor now has one.
- Run the 60-day test: if that model were switched off in two months, what breaks, and how long to move?
- Name one part of your stack that would survive a model swap untouched. That is the part worth funding.
What Not To Do
Don’t treat picking a model as a one-time strategic decision. It is a procurement choice you will re-make on the vendor’s schedule, not yours. Don’t pour budget into chasing this week’s coding leaderboard. The lead changes monthly. And don’t read “we are on the best model” as a moat. This week’s best model is next quarter’s deprecated one.
Signal Boost
Anthropic’s Model Deprecations page lists the live retirement date for every Claude model. Read it before assuming the model you built on this year still answers next year.