Blog

The model behind your coding agent can disappear. Plan for it

The Memmo team · · 5 min read

Most teams pick a coding agent and think of the model as a setting inside it. Two pieces of news this month are a reminder that the model is a supplier, and suppliers change. On 13 November, GitHub announced that six models will leave Copilot on 14 December. And 12 November was the date OpenAI set to stop supplying its models to Cursor.

Neither is a crisis. Both have replacements. But they show something worth planning for: the model behind your team's agent can change on someone else's schedule, and the places where your team quietly depends on a specific model are rarely written down.

What GitHub is retiring

According to the GitHub changelog, these models will be deprecated across all Copilot experiences on 14 December 2026, with the suggested replacement in brackets:

  • Claude Fable 5 (Claude Fable 5.2)
  • Claude Haiku 4.5 (Claude Haiku 5.5)
  • Gemini 3.8 Flash (Gemini 4 Hydrogen)
  • Gemini 4 Argon (Gemini 4 Carbon)
  • Grok 4.6 (Grok 4.7)
  • Kimi K3 (replacement to be announced)

Claude Haiku 4.5 stays available to individual subscribers on annual plans. For Copilot Business and Enterprise, the suggested alternatives are turned on automatically under default model enablement, unless an administrator has switched off the global default or that specific model. In that case, an admin has to enable the replacements in the Copilot model policies. GitHub asks customers to move their workflows and integrations to supported models before the date.

What happened with Cursor

The Cursor case is different, because it is about a business relationship, not an upgrade. As The Next Web reported, OpenAI said on 29 August that it would end its contract to supply models to Cursor following SpaceX's acquisition of the company, with access ending on 12 November. OpenAI said that was the longest notice its contract allowed, and that its next model would not be offered in Cursor at all.

An opinion piece in CIO by architect Vipin Jain uses this as an example of why coding agents should be managed like any other supplier. He notes that Cursor's CEO reportedly put OpenAI's share of Cursor's traffic at about 5%, and that the notice period was 76 days. His advice is simple: assume change.

Why this matters for a team

When a developer picks a model from a menu in their editor, a retirement is a small annoyance. They pick another one. The trouble is everywhere else a model name gets written down and forgotten:

  • a CI job that runs an agent with a fixed model flag;
  • a review bot or internal script that calls an API with a hard-coded model ID;
  • an organization policy that only allows a short list of models;
  • a prompt, rule file or skill tuned over months for how one model behaves;
  • a cost estimate or budget built on one model's price and token use.

Those are the places that break or drift quietly. A replacement model is not a drop-in copy. It can be stronger and still follow your instructions differently, use more or fewer tokens, or cost more per task. The new model may well be better. You just want to find out on your own terms, not on the morning a pipeline fails.

What to do before 14 December

  • Find every model name. Search your repositories, CI configuration and internal tools for model IDs. Ask who runs agents or bots outside the editor. The list is usually longer than people expect.
  • Check your Copilot policies. If your organization turned off default model enablement, the replacements won't appear on their own. An admin needs to enable them before the old models go.
  • Test the replacement on real work. Run a handful of recent tasks with known good results through the new model and compare: did the tests pass, how much did a reviewer have to fix, and what did it cost per task. We described a simple version of this in our piece on choosing between new top models.
  • Decide where to pin and where to float. Interactive work can usually follow the vendor's default. Unattended jobs, like CI or scheduled agents, are better pinned to a named model so they don't change under you, with a calendar reminder to review the pin.
  • Keep a second option ready. If a team relies on one provider for everything, write down which model it would switch to and check, once, that it works with your setup. The Cursor case shows a supplier can leave for reasons that have nothing to do with the technology.

Make the next switch boring

Retirements like this will keep coming. New models now arrive every few weeks, and older ones have to make room. The teams that handle it best will be the ones whose setup does not depend on one model's quirks: rules and project context written in plain language any model can follow, tests that decide whether a change is good, and one person or team who owns the list of agents and models in use.

That is the real takeaway from both stories. The model is the part of your coding setup you control least. Put your effort into the parts you do control, and a model change becomes a configuration change, not a project.

Sources: GitHub changelog, “Upcoming deprecation of selected GitHub Copilot models in December” (13 November 2026), The Next Web on OpenAI ending model supply to Cursor (29 August 2026), Vipin Jain in CIO (9 November 2026).

Give your agents and reviews your team's memory.

One command in your repo. Nothing reaches an agent until a person approves it.

One short guide a week on AI coding agents.

No spam. Privacy.