Copilot Custom Agents Pinned to Retiring Models: Aim at the LTS, Not the Suggested Alternative

On 18 September GitHub announced that six models leave every Copilot surface on 19 October 2026: chat, inline edits, ask and agent modes, and code completions. GPT-5.5 and GPT-5.4 are replaced by GPT-5.6 Sol, GPT-5.4 mini and GPT-5 mini by GPT-5.6 Luna, Gemini 3.7 Flash by Gemini 3.8 Flash, and Grok 4.5 by Grok 4.6. That is 31 days of notice, 27 of which are left today.

For people who pick a model from the chat dropdown, this is a menu change. For Copilot custom agents and prompt files that name a model in their frontmatter, it is a migration, and the obvious target is the wrong one.

Copilot custom agents already pin these models

GitHub code search, run today, finds 1,242 public .agent.md files containing the line model: GPT-5.4 (the phrase match also catches the mini variants), 297 with model: GPT-5 mini, and 100 with model: GPT-5.5. A further 178 .prompt.md files carry model: GPT-5.4. For scale, 30,592 .agent.md files contain model: at all. A sample of twenty GPT-5.4 hits were all real frontmatter pins, in forms ranging from gpt-5.4 to GPT-5.4 (copilot).

Those are file counts on indexed public default branches, not repository counts, and the .agent.md convention is not exclusive to Copilot, so read them as an upper bound on public exposure. Private repositories are invisible to this search.

What happens to a file whose single pinned model is retired is not stated in either reference. GitHub's custom agents configuration page types model as a string and says an unset value inherits the default. The VS Code page says an unset value falls back to whatever the model picker has selected. Neither says what a string naming a retired model does.

The assumption this breaks: the suggested alternative is a stable target

The announcement tells you to move to the suggested alternative, and the docs carry a retirement table built for that purpose. Open its source file, model-deprecation-history.yml in the github/docs repository, and three things show up.

First, it has not caught up. It holds 36 rows and none is dated 19 October, although the file was last edited on 21 September, three days after the announcement. A team that plans migrations from the docs table, rather than the changelog, does not see this retirement at all.

Second, it still sends people to models that are leaving. Ten of the 36 rows name an alternative that is retiring by 19 October: GPT-4.1, GPT-5, GPT-5.2 and o3 all point at GPT-5.5; o1-mini, o3-mini, o4-mini and Grok Code Fast 1 point at GPT-5 mini; and Gemini 3 Flash and Gemini 3.1 Pro point at Gemini 3.6 Flash, which the same file retires on 2 October.

Third, the dates in that file measure how long a suggested alternative lasts. A team that followed the advice when GPT-4.1 retired on 1 June got 140 days on GPT-5.5. GPT-5.2 users got 136. Grok Code Fast 1 users moved to GPT-5 mini on 15 May and got 157 days. Gemini 3.1 Pro users moved to Gemini 3.6 Flash on 1 September and get 31.

The replacement named on 18 September was not the newest in its family for even three days. Grok 4.6 is the suggested alternative for Grok 4.5; on 21 September GitHub began rolling out Grok 4.7, described as building on 4.6.

None of this is a docs team failing. It is what a catalogue that turns over this fast looks like when you read it as a support policy, which it is not.

The one target with a written commitment

The docs page on base and LTS models is the only place that promises a duration. GitHub designated GPT-5.3-Codex as both the base model and the LTS model on 18 March 2026, and commits to keeping an LTS model available for one year from designation, which runs to 18 March 2027. The commitment applies to Copilot Business and Copilot Enterprise only.

Public agent files already lean this way: 746 .agent.md files pin model: GPT-5.3-Codex, against 43 that pin model: GPT-5.6 Sol, the headline replacement.

VS Code gives you the mechanism to use it without giving up a newer model. Its custom agent reference lets model be an ordered list, and the editor tries each entry until it finds one that is available. The GitHub reference for the cloud agent documents only a string, so do not assume the list form works there.

This is the same last-mile problem described in why enterprise AI doesn't ship: the model was never the fragile part. A custom agent is an integration, and an integration pinned to a vendor's release train inherits its churn. Earlier this month the same catalogue showed up in the Copilot auto-model tiers, where the choice of model was already being taken out of the file.

What to do before 19 October

1. Find every pin in your repositories:

grep -rniE '^model: *(gpt-5\.5|gpt-5\.4|gpt-5[ -]mini|gemini[ -]3\.7[ -]flash|grok[ -]4\.5)' .github

Across an organization, a broader code search through the API lists the candidate files to open and check:

gh api -X GET search/code -f q='model GPT-5.4 org:YOUR_ORG filename:.agent.md' --jq '.items[].html_url'

1. For agents that run in VS Code, replace the single pin with a list whose last entry is the LTS model, for example model: ['GPT-5.6 Sol (copilot)', 'GPT-5.3-Codex (copilot)'].
2. For agents that target the cloud agent, pin GPT-5.3-Codex or remove the field so it inherits the default. Any other pin is a date you will be editing again.
3. Put 18 March 2027 in the calendar. That is when the LTS commitment on GPT-5.3-Codex runs out, and the next designation should be read the day it lands.

What would make this wrong

The advice is wrong if GitHub designates GPT-5.6 Sol or Luna as the next LTS model before March, since then the suggested alternative and the committed one are the same. It is also wrong for anyone on Copilot Pro or Pro+, to whom the LTS commitment does not apply. And if GitHub documents list support for the cloud agent, step 3 collapses into step 2.