On 16 November, GitHub retired three Copilot models: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. The changelog names their replacements and adds a line that sounds like good news: no action is required to remove the deprecated models.
Read that line as an engineer and it gets less comfortable. If any step of your work ran on one of those models, it now runs on something else or doesn't run at all, and I'd bet nothing in your setup told you. A model is a dependency you don't control, so treat it like one: write down which model each AI step uses, add a small check that fails loudly when one changes, and name a fallback before you need it.
What the changelog actually says
I only read GitHub's changelog entry for this, so that's all I'll claim. Here's the useful part, without the noise:
- Replacements: Claude Opus 5.5 for both Claude Opus 4.8 and Claude Opus 5. GPT-6.1 Sol or GPT-6 Luna for GPT-5.6 Terra.
- Where it applies: across most Copilot experiences, which means Copilot Chat, inline edits, ask and agent modes, and code completions.
- Notice: it was announced on 15 October, about a month before it took effect.
- Two exceptions: individual subscribers on annual plans can still use Claude Opus 4.8. Claude Opus 5 leaves the model picker but stays available for runtime refusal fallback through existing authorized access.
- For Business and Enterprise: the suggested alternatives are switched on automatically under default model enablement, unless an admin turned off the global default or disabled the model. If the global default is off, admins enable them through model policies in Copilot settings. Existing policies keep applying.
- What you're asked to do: update your workflows and integrations to supported models.
That last line is the whole article in one sentence from the vendor. Update your workflows. Which assumes you know which of your workflows use which model.
Why "no action required" is the line to worry about
We would never accept this for a library. Picture a package in your build quietly swapped for a newer major version overnight, with a note saying you don't need to do anything. You'd want a lockfile, a test run and a way back.
A model sits in the same place in your workflow. It writes your completions, drafts your refactors in agent mode, maybe summarises tickets inside your product. When it changes, the shape of what it writes changes too, even if every request still succeeds. Green, and different.
Before
The model as a setting
- Someone picked it in a dropdown months ago
- Nobody can say which steps depend on it
- A swap shows up as "the suggestions feel off lately"
After
The model as a dependency
- Every AI step has its model written down next to it
- A check fails the day that model changes
- The fallback is already named, so the swap is a decision, not a surprise
Write down the model under every step
Start with a plain file in the repo. Nothing clever. One line per place where a model touches your work, the model it uses today, and the one you'd move to.
- 1List the steps, not the tools"Copilot" is not an answer. Completions, chat, agent mode and any AI feature you ship to users are separate steps, and they can sit on separate models.
- 2Write the exact model nameNot "the Claude one". The name you'd find in a changelog, so a deprecation notice can be matched against your list in a minute.
- 3Add the fallback next to itPick it now, from the replacements your vendor names or from what you've already tried. Choosing calmly beats choosing on the morning something stops working.
A check that fails loudly
The list is only useful if something compares it with reality. Where your own code or config picks the model, a tiny check can read that value and compare it with the pinned one. When they differ, it fails, in CI, in red, where a person sees it.
The failure isn't the point. The point is that the change becomes a commit with your name on it, and that it triggers the few saved prompts you run to see whether the new model still writes the way your code needs.
Name the fallback while you're calm
GitHub gave about a month of notice. A month is plenty if a deprecation notice lands on a list you already keep. It's nothing if the first time you hear about it is a teammate asking why agent mode got chattier.
If you can't say which model writes your code today, you won't notice the day it stops being that one.
So this week, do three small things: write the list, add the check, pick the fallbacks. It's a small job, and the next deprecation becomes a line in a diff instead of a mood in the team.
Free resourceThe Unattended AI Runbook25 checks for AI that runs while you sleep. A good companion if your models now sit in a list you actually check.Get the runbook


