Pimp My IDE / garage dispatch
Back to the garage
October 6, 2026 | Copilot / model routing / migration

The model picker is not the whole circuit.

GitHub Copilot will retire six models on October 19. A replacement in the picker does not repair a pinned custom agent, a scheduled task, or a background utility model.

Inventory every model route before the cutoff. Then replay the work.
19OCT 2026

6 MODEL ROUTES SCHEDULED TO OPEN
13 DAYS FROM THIS SOURCE REVIEW

Six routes have a date.

GitHub says Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 will leave all Copilot experiences on October 19, 2026. The notice names a suggested alternative for each model. Enterprise and Business accounts using default model enablement receive those alternatives automatically unless an administrator changed the global default or disabled a model.[1]

That policy change makes a replacement available. It does not prove that the replacement behaves the same, costs the same, or works in every client your team uses. GitHub's supported-model page says model availability depends on the plan and client. It also lists minimum IDE versions for recent models and warns that availability can change.[2]

A suggested alternative restores a route. It does not certify the workload.

Model state hides outside the picker.

VS Code custom agents can pin one model or a prioritized model list in an .agent.md file. A list gives the system an ordered fallback, but only among models that are available. User-level agents run from the selected Agent Host, not from VS Code profile storage.[3]

VS Code also keeps model choices in separate places. Chat has a model picker. Inline chat can use inlineChat.defaultModel. Inline suggestions have their own completion model. General and small background tasks can use chat.utilityModel and chat.utilitySmallModel. Auto selection can route by task complexity and current availability.[4]

A migration that checks only the visible chat model can miss commit messages, pull request text, intent detection, inline suggestions, repository agents, user agents, and scheduled work. Search the files and settings that own those routes.

Do not treat fallback as parity.

A new model can change patch shape, tool choice, latency, credit use, refusal behavior, and how much context it consumes. A fallback list keeps a task from stopping when one model is unavailable. It does not prove that the second model meets the first model's acceptance bar.

Use a matched replay. Keep the task, repository revision, harness, tools, permissions, context cap, and review method fixed. Record the requested model and the model that served the run. Compare the patch, checks, review time, tool calls, wall time, and credit use.

Run the retirement drill before October 19.

  1. Search repository and user agent files for every retiring model name.
  2. Inspect chat, inline chat, completion, utility, Auto, and organization model settings.
  3. List scheduled prompts and unattended tasks that depend on a model route.
  4. Enable the replacement where policy requires an explicit choice.
  5. Replay known work with the replacement and record the served model.
  6. Test the unavailable-model path. Confirm that the task stops clearly or uses an approved fallback.
Interactive makeover / route inventory

Model retirement fuseboard

Traditional purpose replaced: switch one model in one picker. Better version: choose the retiring route, close each place that can hold model state, and copy a migration card that still requires a real replay.

Choose the retiring route
Close the route inventory

These controls write an inventory template. They do not inspect your settings, change a model policy, or run a replay.

Gemini 3.7 Flash to Gemini 3.8 Flash2 of 5 circuits closed

The visible routes are listed. Utility, automation, and replay circuits remain open.

Gemini 3.7 Flash is selected. 2 of 5 inventory circuits are closed.

Sources read

Source log and evidence boundary
  1. GitHub Changelog, "Upcoming deprecation of selected GitHub Copilot models in mid-October", September 18, 2026, read October 6, 2026. It names the six models, October 19 cutoff, suggested alternatives, scope, and default-enablement behavior.
  2. GitHub Docs, "Supported AI models in GitHub Copilot", read October 6, 2026. It documents plan and client differences, minimum IDE versions, model policy, and retirement history.
  3. Visual Studio Code Docs, "Custom agents in VS Code", updated September 30, 2026, read October 6, 2026. It documents single-model pins, ordered model lists, agent file locations, and Agent Host storage.
  4. Visual Studio Code Docs, "AI language models in VS Code", read October 6, 2026. It documents chat, Auto, inline chat, completion, utility, and bring-your-own-key model routes.

Evidence boundary. We read GitHub and VS Code documentation. We did not inspect a private organization's model policies or run paid Copilot replays. GitHub's alternatives are recommendations, not evidence of workload parity. The fuseboard writes a migration template. Approval still needs real settings, a known workload, a served-model record, and reviewed output.