If your agent mode still pins Claude Sonnet 4.6 with a sticky note that says “migrate later,” the calendar is no longer abstract.
Core takeaway: On 2026-09-01, GitHub removes six models from every Copilot surface — Chat, inline edits, ask and agent modes, and code completions — and on 2026-08-26 the Business/Enterprise Default availability for released models policy starts actually governing unconfigured GA models. A model pin is a runtime dependency, not a preference.
Why it matters
This is not a cosmetic rename in the model picker. GitHub’s 2026-07-31 changelog states the blast radius explicitly: Copilot Chat, inline edits, ask and agent modes, and code completions. If a CI job, Copilot agent policy, or onboarding doc hard-codes one of the six IDs, that string stops being a valid route on September 1.
There is one narrow exception: Claude Sonnet 4.6 remains available to individual Copilot subscribers on annual plans. Monthly individual plans and org/enterprise seats do not get the carve-out. “Sonnet 4.6 is our house style” works until half the team is on Business seats and the picker goes empty on Monday morning.
A second clock is easy to miss. The Default availability for released models policy (announced 2026-07-29) was configurable but inert for ~28 days; from 2026-08-26 unconfigured GA models inherit the policy’s enabled/disabled default. Explicit enable/disable choices stay put. Open-weight models and models outside GitHub’s data-retention agreement (e.g. Fable 5) stay out of default enablement. Deprecation cleanup and auto-enable governance land in the same window — treat them as one ops story.
How it compares
| Event | When | Character |
|---|---|---|
| Gemini 2.5 Pro / Gemini 3 Flash retired | 2026-07-31 (immediate) | Already landed — alternatives 3.1 Pro (Preview) / 3.6 Flash |
| Six-model notice retirement | 2026-09-01 | Chat + agents + completions |
| Default availability takes effect | 2026-08-26 | Business/Enterprise; only unconfigured GA models follow the default |
| Enterprise teams model targeting (preview) | ~2026-08-03 opt-in for most | Team-scoped grants; with teams mode on, org-level model settings no longer apply |
One sharp comparison: the Jul 31 Gemini note steered 2.5 Pro users toward Gemini 3.1 Pro (Preview), while the same-day Sep 1 table retires Gemini 3.1 Pro in favor of Gemini 3.6 Flash. Gemini shops that “just upgraded” once may need a second hop — prefer 3.6 Flash as the durable default unless you have a deliberate Preview reason.
Also: Cursor Router (pick a model per turn in the IDE) is not Copilot tenant policy (which model IDs exist for your org). Calling both “auto” is how audits get skipped.
Code / config example
Official migration table
| Retiring (2026-09-01) | Suggested alternative |
|---|---|
| Gemini 3.1 Pro | Gemini 3.6 Flash |
| Claude Opus 4.5 | Claude Opus 4.7 / 4.8 / 5 |
| Claude Opus 4.6 | Claude Opus 4.7 / 4.8 / 5 |
| Claude Sonnet 4.5 | Claude Sonnet 5 |
| Claude Sonnet 4.6* | Claude Sonnet 5 |
| Raptor Mini | MAI-Code-1-Flash |
*Individual annual-plan exception only.
Pin-audit sketch
A starting point for sweeping hardcoded model strings out of repos and runbooks (paths vary by org):
# Example workspace sweep for retiring identifiers
rg -n --hidden -g '!node_modules' -g '!.git' \
-e 'claude-sonnet-4\.6|claude-opus-4\.(5|6)|gemini-3\.1-pro|raptor-mini|Raptor Mini' \
.
# Then track: path | current | successor | owner | retested?
Enterprise admin path (docs): Enterprise settings → AI controls → Copilot → Configure models. Set Enabled / Disabled / Optional per model, and review Default availability for released models. Successors will not appear in the picker until policy allows them.
[ ] Search policies/runbooks/agent YAML for the six Sep 1 models
[ ] Enable suggested successors in Configure models (or Optional → org enable)
[ ] Confirm alternatives in VS Code + github.com Copilot Chat pickers
[ ] Smoke agent mode, inline edit, and completions separately
[ ] Before Aug 26: keep default-availability enabled vs disable for manual approve
[ ] If using Enterprise teams mode (preview): document that org model settings stop applying
Practical use
Individuals (Pro and similar): Flip the picker, then spend two days noticing whether your custom instructions assumed a particular verbosity or tool-calling style. Leaning on the Sonnet 4.6 annual carve-out can desync you from teammates on org seats.
Business / Enterprise admins: GitHub’s “update your workflows and integrations” is an audit, not a one-line diff. Include Copilot Extensions, agent policies, “our default model” wiki pages, and onboarding decks. Successor models often need an explicit Enable before anyone sees them — Enable → verify picker → then tell the team.
Quality regression: Review prompts tuned on Opus 4.6 may ramble or tool-call differently on Opus 5 / Sonnet 5. Skip the public leaderboard argument; A/B three representative PRs from your own monorepo.
I once assumed “if it’s in the docs table, the picker will just show it,” skipped Configure models, and watched a staging agent die with a missing-model failure. Since then the runbook gets an Enable screenshot, not just a changelog link.
Senior-engineer perspective
Treat model pins like exact versions in a lockfile — except the vendor owns the EOL date.
- Explicit pins buy reproducibility and bind you to the provider calendar.
- Auto / default routing reduces ticket noise; without telemetry you still get silent quality drift.
- The Aug 26 default-availability switch is governance: “ship GA models to users by default” vs “compliance approves each one.” Remember open-weight and non-retention models stay outside default enablement so you do not hallucinate a total auto-on world.
Using it in Cursor
If you mix Cursor agents with Copilot (PapaCoder Labs does):
- Keep Cursor Router / model settings and Copilot pins in separate runbook sections.
- Only repos that actually invoke Copilot agent mode belong in the Sep 1 audit set.
- Sweep Cursor automation prompts too if anyone hardcoded a Copilot model id (rare, but cheap to search).
- Loop: enable successor → smoke ask/agent on a tiny PR → human reads the diff before merge/publish.
A useful Cursor follow-up prompt: “Run rg for retiring model strings, fill a path/successor/retested table in RUN_REPORT.md, and do not publish.”
FAQ
Q. Do we remove old models from config after September 1?
A. GitHub says no cleanup is required on their side — they stop offering them. Your pins, scripts, and docs are still your mess to delete.
Q. Does the Sonnet 4.6 exception cover our org?
A. No. It is for individual annual subscribers. Business/Enterprise lose Sonnet 4.6 on Sep 1.
Q. What if we disable default availability before Aug 26?
A. Unconfigured GA models will not auto-enable. Explicit enable/disable rows stay as set. Choose that if you want manual approval governance.
Q. Is Raptor Mini → MAI-Code-1-Flash a drop-in?
A. It is GitHub’s suggested alternative. MAI-Code-1-Flash is Microsoft’s small-tier coding model tuned for Copilot harnesses. Retest inline/light tasks; do not assume identical quality.
Q. What changes if we opt into Enterprise teams model targeting?
A. Access is evaluated at enterprise + team scope; org-level model settings stop applying. Preview allows rollback — prepare team assignments before flipping the toggle.
Sources
- Upcoming August 2026 model deprecations in GitHub Copilot (2026-07-31)
- Gemini 2.5 Pro and Gemini 3 Flash deprecated (2026-07-31)
- Default model enablement for Copilot Business and Enterprise (2026-07-29)
- Enterprise teams model policy targeting in public preview (2026-07-31)
- About default availability of Copilot models
- Managing availability of Copilot models
- MAI-Code-1-Flash is now available for GitHub Copilot (2026-06-02)
Closing
It reads like model-market news; operationally it is a dependency expiry. Put Aug 26 (default availability) and Sep 1 (six retirements) on the same runbook line. For a while, the expensive Copilot change will be less “new model in the blog” and more one Enable row in AI controls.