Since 27 July: two Gemini models deprecated, GitHub Spark retired, GPT-5.6 Luna cut by eighty percent, four model vendors in one Copilot subscription, and three new ways for an agent to reach into your systems.
A reasonable question after a fortnight like that is what is actually safe to commit to.
Do not standardise on a model
Gemini 2.5 Pro was a flagship. Gemini 3 Flash is newer than that. Both were named for deprecation on 31 July.
At current cadence a model has roughly a year. Anything you write down that names a specific model — a prompt tuned to its quirks, an evaluation baseline, a documented team standard — has that same shelf life.
Do not standardise on a platform you cannot leave
GitHub retired Spark on 3 August. Prompt-to-app builders are the most crowded segment in this directory and the least differentiated, because they all sit on the same models.
The test is simple: can you export the code and run it without the platform? If not, you have not chosen a tool, you have chosen a dependency.
Do standardise on these
These are the decisions that survived the fortnight intact.
- A review process that assumes the author is an agent — dependency diff first, requirement re-read, tests before implementation
- Small, reviewable changes; GitHub put stacked pull requests in public preview on 30 July for exactly this reason
- Model access as configuration, so switching vendor is a settings change rather than a project
- One owner for model policy, since it now both grants and removes capability
- Security coverage that scales automatically — code scanning across all repositories, Dependabot alerts that reach a person
The one-line version
Standardise on how you work, not on what you work with. Everything in the second category was replaced at least once in the last ten days.