GitHub Copilot September 2026: Model Deprecations, Gemini 3.8 and Signup Changes

September 4, 2026 GitHub Copilot update: GitHub’s September 3 changelog says selected models are deprecated across Copilot experiences on October 2, 2026: Gemini 3.5 Flash → Gemini 3.8 Flash, Gemini 3.6 Flash → Gemini 3.8 Flash, Kimi K2.7 Code → Kimi K3, and Claude Opus 4.7 → Claude Opus 5. GitHub says no removal action is required, but administrators may need to enable alternatives and update workflows. Gemini 3.8 Flash is in gradual rollout across listed plans and surfaces. Business and Enterprise signups are gradually reopening with documented seat and billing timing; verify organization state before acting.

September 23, 2026 update: GitHub’s September 18 changelog schedules six additional Copilot models for deprecation on October 19. The model mapping and Business/Enterprise policy conditions are in the existing model-deprecation section below.

September 3, 2026 update: Two September 2 GitHub Changelog entries add governance controls to the existing model, budget and review checklist. Enterprise-managed settings can support an administrator-selected default model, and content exclusions are generally available in the Copilot app and CLI. The controls have different surfaces and limits, so verify the product and policy scope before treating either as universal.

Source date: GitHub published the VS Code Agents usage update and code-review update on September 11, 2026. Availability still depends on plan, policy, repository and product surface.

What changed in the September 11 update

GitHub’s usage-metrics entry adds fields for VS Code Agents, including daily active users and totals by the dedicated VS Code agent surface. GitHub distinguishes that surface from editor Agent Mode and generic Copilot rollups. A blank or null value can therefore mean that the feature is unavailable or that the field does not apply; it should not be silently converted into zero usage.

A separate code-review entry describes addressed-comment auto-resolution, smart commit messages and analysis changes. GitHub also describes an agent ensemble in Lite and broader shell-tool work behind firewall controls. The percentages in the announcement are GitHub’s own experiment results. Keep them labelled as vendor evidence, not as an independent benchmark or a promise that a repository will see the same outcome.

A decision matrix for Copilot administrators

A safe admin checklist

1. Identify the surface. Record whether the question concerns VS Code Agents, editor Agent Mode, the Copilot app, CLI, JetBrains or code review. Do not compare unlike counters.
2. Record the policy boundary. Note organisation, enterprise, repository and user settings that can control access. A changelog entry is not a live entitlement read-back.
3. Choose a review-only pilot. Select a low-risk repository, keep branch protection and required human review in force, and record the commit SHA and feature setting.
4. Inspect the proposed change. An addressed comment, smart commit message or analysis result still needs a human owner to validate scope, tests, security and attribution.
5. Measure the right denominator. Keep dedicated VS Code Agent fields separate from other Copilot usage. Document null handling and the time window before comparing reports.
6. Write a rollback rule. Disable the experiment or revert the workflow if review quality, policy compliance or repository safety is not acceptable.

What the weekly release does not prove

The September 7 weekly release lists Jira integration in the Copilot app, HydraFusion experimental CLI routing, VS Code automations, voice and issue work, and a managed JetBrains sandbox. These descriptions are useful discovery pointers. They do not prove a common availability date, identical permissions or production readiness across surfaces. Verify the current documentation and account state before changing a team workflow.

September 2 additions: default models and content exclusions

Default model policy

GitHub’s default-model changelog says enterprise-managed settings support any preferred default model for new conversations, with team overrides through the documented model and team-mapping controls. The relevant administrator task is to record the intended default, the teams allowed to override it and the review owner. “Available in the setting” does not mean every model is available to every plan, organisation, user or surface.

Content-exclusion policy

GitHub also says content exclusions are generally available in the Copilot app and CLI and that the app/CLI respect enterprise, organisation and repository policies. The Changelog entry should be read with the content-exclusion documentation: an administrator configures paths or fnmatch-style patterns, propagation can take up to 30 minutes, and the setting must be tested. The current documentation notes a limitation for Agent mode in IDEs, so do not claim that one exclusion configuration blocks every Copilot surface.

Use GitHub’s enterprise-managed-settings reference as the schema authority and its configuration guide for the admin path. Test one repository with a known excluded path, one included path and the actual client surfaces your teams use. Save the observed policy state and timestamp; do not infer a successful exclusion from a setting that has not propagated. The prior model-deprecation, multi-organisation access, individual-budget and code-review sections below remain part of the same reader job and should not be removed or shortened.

Secondary Changelog coverage and community discussions help identify administrator questions, but they do not establish plan entitlement or enforcement. For adjacent implementation, use the Copilot VS Code release guide for feature changes, the agent/plugin guide for tool governance and the agent-harness guide for state and evaluation. This page remains the September governance canonical.

The two September 2 entries are documented product-policy updates. They do not prove that a particular organisation has the setting enabled, that a user can select every model or that an exclusion covers every client. Confirm account-level readback before changing production policy.

Short answer: GitHub’s September 2026 Copilot changes require an admin inventory: map deprecated models, verify the billing organisation, set budget expiry deliberately and pilot code-review approvals before relying on them.

Four GitHub Changelog entries from 31 August and 1 September 2026 change Copilot governance rather than adding one isolated feature. This checklist turns the model, access, budget and review-policy changes into explicit owner decisions.

The September admin checklist

GitHub’s 31 August and 1 September 2026 Changelog entries form one practical administrator update. The changes touch four controls: selected model deprecations, Team-plan model access when a user belongs to multiple organisations, individual-user budget expiry, and Copilot code-review approvals.

The safest response is an inventory, not a panic migration. Check which repositories and teams depend on a named model, identify the billing organisation for multi-seat users, record budget owners and expiry dates, and decide whether code-review approval belongs in branch protection. GitHub says no action is required to remove the deprecated models, but that does not remove the need to test the replacement workflow.

This page complements the GitHub Copilot VS Code August release guide, which owns feature changes. It does not absorb the model-selection question covered in the existing Fable versus Sonnet comparison.

Model deprecations: map by family, then test

October 19, 2026: six additional model deprecations

GitHub’s September 18 notice schedules the following models for deprecation across Copilot Chat, inline edits, ask and agent modes, and code completions on October 19. The right column is GitHub’s suggested alternative, not a claim that every workflow will behave identically after a switch.

Update any workflow or integration that explicitly selects one of the retiring models before October 19, then test the suggested alternative on representative tasks. GitHub says no action is needed to remove the old model after deprecation; that does not verify that a replacement is enabled for your users or that a pinned workflow has moved to it.

Business and Enterprise policy check: GitHub says the suggested alternatives are automatically enabled under the default model enablement policy unless an administrator has turned off the global default or explicitly disabled a model. If the global default is off, an administrator can enable access through model policies. After a model is enabled, users can select it in supported Copilot experiences. For individual plans, and for every organization or client surface, verify the current plan and supported-model list; a suggested alternative is not proof that it appears for every seat or IDE.

There is a documentation timing gap: as checked September 23, the supported-models page still lists the affected models as supported before their scheduled cutoff, while its retirement-history table shows the earlier October 2 retirements but not this announced October 19 schedule. The September 18 changelog is the date authority; re-check it and your model settings before the cutoff.

GitHub’s 31 August Changelog says selected Copilot models were deprecated from 1 September 2026 and lists replacements: Gemini3.1 Pro to Gemini3.7 Flash, Claude Opus4.5/4.6 to Opus4.7/4.8/5, Claude Sonnet4.5/4.6 to Sonnet5, and Raptor Mini to MAI-Code-1.1-Flash. Use that list as a lifecycle map, not as a claim that every replacement behaves identically.

September 10 addition: GitHub’s September 10 changelog lists MAI-Code-1-Flash as deprecated and names MAI-Code-1.1-Flash as the suggested replacement. Treat that as a lifecycle mapping, not a parity guarantee: verify availability, context behaviour, tools, latency, quality and cost in the organisation’s own plan before changing a default or automation.

GitHub says no action is required to remove the deprecated models and administrators may enable alternatives. Still, teams should test the prompts that matter: code generation, review comments, tool calls, long-context tasks and any marketing-operations automation that consumes the output. Record the model, repository, prompt class and observed failure mode.

Do not update a production runbook solely by replacing a string. A model deprecation can change latency, context behavior, tool permissions, quality and cost. Keep the rollback or fallback decision with the repository owner and verify current policy in the live organisation.

• Inventory model names in prompts, policies, docs and automation.
• Map every deprecated family to GitHub’s listed replacement family.
• Run representative tests with the replacement before changing a production default.
• Record model, date, repository, task class and reviewer.
• Keep a human fallback for critical workflows.

Team access now follows the billing organisation

GitHub says that when a user has seats in multiple organisations, Team-plan model access is determined only by the billing organisation shown under Usage billed to. That rule explains why a user can have a seat in one organisation but not see the expected Team-plan model access from another.

The operational check is straightforward: identify the seat, open the billing or usage relationship, find the organisation that is billed, and inspect its model policy. Do not ask a user to change seats or permissions until the billing path is confirmed. GitHub says enterprise-only access is unaffected by this specific change.

For an adjacent access-control mindset, use the agent plugins, Skills and MCP guide and the scheduled-workflow permissions guide; those pages are not GitHub documentation, but they reinforce explicit ownership and permission checks.

Individual budget expiry is a governance control

GitHub’s 1 September update lets Business and Enterprise administrators set an expiration date for individual user budgets. The choices are no expiration, the next billing cycle or a specific date; GitHub also exposes an API field named expires_at. The value is operational: a budget can have a review date instead of remaining open-ended by accident.

Assign an owner and reason to each expiry. A temporary project may use a fixed date, while a recurring team budget may align to the next billing cycle. Record what happens at expiry, how the user requests renewal, and whether repository or organisation budgets remain unaffected. Do not assume a budget setting is a complete cost-control system.

Review spend and task value together. A lower cap can protect a budget while also interrupting a critical agent workflow. Set an escalation path before a user reaches the limit and make the renewal decision visible to the technical and business owner.

Copilot code-review approvals are off by default

GitHub says Copilot code review now produces approval assessments in every review, while actual approvals are off by default and can be configured at the enterprise, organisation or repository level. If enabled, the approval can count toward required approvals. GitHub also says a new commit dismisses the approval.

That makes this a branch-protection decision, not a simple quality badge. Before enabling it, define which repositories qualify, what human review remains required, how CODEOWNERS interacts with the setting, and what happens when a new commit arrives. Test the setting in a non-critical repository and keep an audit trail of policy changes.

For wider agent permission context, connect to the Claude Code Auto Mode guide; the product is different, but the governance principle is shared: make machine authority explicit and reviewable.

A practical admin runbook

Start with a change register dated 1 September 2026. For each team, list deprecated models, replacement tests, billing organisation, budget expiry, code-review policy and owner. Add the source link and the evidence date so the register can be refreshed when GitHub changes the product.

Then run the smallest safe validation: one representative task per model family, one multi-seat access check, one budget expiry read-back and one review-policy test in a non-critical repository. No private account or UI check was performed here; these are the questions an administrator should answer before changing policy.

Finally, publish internal guidance that distinguishes what GitHub changed from what your organisation decided. That separation prevents a vendor default, an entitlement condition and an internal control from being reported as the same thing.

Recommended workflow

1. Create a dated inventory of affected models, organisations, budgets and repositories.
2. Map deprecated model families and test listed replacements on representative tasks.
3. Verify the billing organisation for any user with multiple Team-plan seats.
4. Set budget expiry with an owner, renewal path and escalation rule.
5. Pilot code-review approvals in a non-critical repository and re-check branch protection after new commits.

Related Digital Marketer guides

• GitHub Copilot in VS Code August 2026 — Existing Copilot feature-release owner; this page is the administrator-governance owner.
• OpenAI agent plugins, Skills and MCP — Agent integration and control-plane context.
• ChatGPT Work scheduled tasks — Scheduled-workflow governance context.
• ChatGPT Site Tools and WebMCP — Browser and site-tool workflow context.
• Claude Code Auto Mode — Permission and classifier context for agentic work.
• Claude Tag for Marketing Teams — Slack-based Claude workflows, memory and spend controls.

Questions marketers are asking

Do I need to remove deprecated Copilot models?

GitHub says no action is required to remove them and administrators may enable alternatives. Test and document replacements if a workflow depends on a deprecated family.

Why can a multi-organisation Team user lose model access?

GitHub says access is determined by the billing organisation under Usage billed to for this case.

Can an individual Copilot budget expire?

Business and Enterprise admins can choose no expiry, the next billing cycle or a specific date, according to GitHub.

Can Copilot approve a pull request automatically?

GitHub says approval assessments are available, actual approvals are off by default and configurable. If enabled, they can count toward required approvals; new commits dismiss the approval.

Sources and dates

Publication, event and update dates are kept separate below. The September 18 model notice and current model-policy documentation were accessed September 23, 2026; older entries retain their original access dates.

• GitHub Changelog: upcoming deprecation of selected GitHub Copilot models in mid-October — published September 18, 2026; lists six models scheduled for October 19 and the suggested alternatives and Business/Enterprise default-enable policy conditions; accessed September 23, 2026.
• GitHub Docs: supported AI models in GitHub Copilot — checked September 23, 2026 for plan/client availability and retirement-history coverage. Its current model tables still listed affected models as supported before cutoff; the retirement-history table showed the October 2 wave but not the newer October 19 schedule.
• GitHub Docs: default availability of Copilot models and configure access to Copilot models — checked September 23, 2026 for Business/Enterprise policy scope and the effect of explicit model settings.
• GitHub Changelog: selected Copilot models deprecated — published/event date: 2026-08-31; accessed 2 September 2026. GitHub lists model replacements effective 1 September 2026 and says no removal action is required; administrators may enable alternatives.
• GitHub Changelog: Copilot Team model access update — published/event date: 2026-08-31; accessed 2 September 2026. For users with seats in multiple organisations, GitHub says access is determined by the billing organisation under Usage billed to.
• GitHub Changelog: expiration dates for individual user budgets — published/event date: 2026-09-01; accessed 2 September 2026. GitHub says Business and Enterprise admins can set no expiry, next billing cycle or a specific date through the API or admin controls.
• GitHub Changelog: Copilot code review approvals — published/event date: 2026-09-01; accessed 2 September 2026. GitHub says approval assessments are available, actual approvals are off by default and can be configured; approved reviews count toward required approvals when enabled.

Re-check platform documentation immediately before implementation. Product rollouts, eligibility, pricing, policy, and measurement definitions can change after publication.

Sources: GitHub Changelog: MAI-Code-1-Flash deprecated