GitHub Copilot Weekly Releases — September 21, 2026: What Changed

Which AI models were added to GitHub Copilot in the September 21 week?

GitHub's roundup groups four model announcements. They do not share identical plan eligibility. The following table reflects what GitHub reported for that launch week, not a live list of models that every account can select on November 8.

GitHub published a separate Opus 5.5 release note describing early internal task results and gradual rollout. These are GitHub's observations, not independent SXF benchmarks. A model appearing in an announcement does not prove that it is cheaper or better than competitors on a particular repository.

Availability caveat: a model's current presence in the picker depends on your Copilot plan, client and organization or enterprise settings. GitHub's supported-models reference is the correct source for today's availability, while its pricing reference explains AI Credits and token rates. Do not assume a model is free merely because it is selectable, or use September pricing as a November invoice estimate.

What should a developer test when switching models?

Use the same issue, repository snapshot and acceptance tests to compare two supported models. Record the patch's correctness, unit-test results, time to verified completion, human review corrections and total Copilot usage. A few attractive generated lines of code do not establish a reliable engineering outcome. The SXF Copilot/Claude setup guide covers selection mechanics; this story focuses on why the September release week changed the choice set.

GitHub Copilot local sandboxing: what did the September preview actually restrict?

GitHub announced local sandboxing for the Copilot app on September 23, 2026. At launch it was a public preview designed to constrain commands initiated by an agent. Project-level settings covered extra read/write paths, read-only or denied directories, outbound and local-network access, Git credentials and GitHub CLI credentials. Enterprise settings could impose stricter effective restrictions. These are execution controls, not model intelligence improvements. Original preview announcement.

At the September preview stage, sandboxing was off by default for the Copilot app, applied to new sessions when enabled at project level, and did not cover cloud-sandbox or remote-host sessions. The preview also documented failure when an operating system could not enforce the requested sandbox. Teams should check the latest client-specific documentation because scope and setup evolved after launch.

SXF security assessment: sandboxing reduces blast radius; it does not verify intent

A file restriction can stop an agent from writing to an unauthorized path, but it does not prove that the agent's proposed edit is correct. Network and credential restrictions reduce exposure; they do not replace branch protection, secret scanning, code review or a real test suite. Run dangerous tasks with least privilege, keep sensitive repositories and credentials out of unnecessary context, and do not give an agent broad filesystem access simply to resolve a tool failure.

OpenTelemetry in the Copilot app: why agent traces matter

GitHub added OpenTelemetry (OTel) configuration through enterprise-managed settings in a September 22 release. Administrators could export agent activity into compatible monitoring tools, including the sequence of requests to models and the tools the agent called. GitHub's documented setup uses the telemetry property in managed-settings.json and an export endpoint. GitHub's OTel announcement.

GitHub says prompt and response content is excluded by default from those exports. Teams should still review any capture settings and where telemetry is sent; metadata about repository workflows can itself have operational sensitivity. Capturing traces also does not establish that every action was correct.

SXF's interpretation: the pairing of sandbox controls and agent traces is more important than either feature in isolation. One asks “What is the agent allowed to do?” and the other asks “What happened during execution?” A production-ready workflow also needs an answer to “Did the tested outcome match the requested change?” That last question belongs to verification and code review, not telemetry alone. Our AI Agent Reliability research explains how to test repeated success, robustness and failure recovery.

GitHub Copilot in Slack and Microsoft Teams: more context, fewer duplicate issues

The September 25 integration update made Copilot more useful where a ticket often begins: in a team discussion. GitHub described the ability to switch models during a conversation, reuse the selected model for the rest of the thread, detect similar issues before creating new ones, and return links connecting created work to its originating conversation.

GitHub's detailed Slack and Teams release note explains the supported context and reliability improvements. Do not assume every message, private file or repository is automatically included: integration authorization and conversation permissions still matter.

SXF analysis: conversation-to-code is a provenance problem

There is real value in linking a GitHub issue to the message where a bug was reported. The danger is mistaking a fast conversion for an approved requirement. Teams should still require a named issue owner, an explicit acceptance criterion, a repository check and a review of any generated ticket. Duplicate-issue detection should be treated as a suggestion, not a guaranteed deduplication system.

What's new for GitHub Copilot in JetBrains 1.18.0?

GitHub's September 22 JetBrains 1.18.0 update covered more than model choice. Assisted approvals, then in public preview, could automatically approve selected low-risk tool calls and still ask for confirmation on higher-risk actions. Editing an earlier user message rewound both conversation and file changes before continuing, giving developers a cleaner way to redirect an agent session. JetBrains release notes.

The update also brought organization- and enterprise-shared skills and managed instructions into local and Copilot agent sessions, a planning mode for the Codex agent, and controls for built-in GitHub MCP tools. For a team, those features create a policy surface: review what instructions are shared, which tools are exposed and which approvals can be automated.

SXF's view: assisted approvals can reduce interruptions, but “low risk” is contextual. A terminal command may be routine in a disposable test environment yet dangerous against production credentials. Treat approval policy and tool scopes as deployment configuration, not as proof that an agent understands consequences.

VS Code 1.139 Copilot update: remote Dev Containers and agent session management

VS Code 1.139 expanded Dev Container agent sessions from local folders to projects hosted on SSH, Tunnel and WSL. The remote host needs Docker and a supported Dev Container configuration. The related setting, chat.agentHost.devContainer.enabled, was gradually rolling out; it might not be enabled by default. The release also improved loading for large agent-session lists and added compact and organizational controls. Official VS Code 1.139 release notes.

For developers with Linux build environments or remote development workstations, executing agent tasks inside the project's actual toolchain can reduce a common source of false confidence: code that passes on the developer's laptop but fails in the target environment. A Dev Container, however, is not automatically a comprehensive security sandbox. Its privileges, mounts, credentials and network access still require review.

For the broader monthly view across VS Code 1.136–1.140, see SXF's September 2026 VS Code Copilot analysis.

SXF analysis: the real September story is governance across five product surfaces

The most important point is not the raw number of model launches. In one week GitHub extended model choice, local agent boundaries, centralized observability, collaboration context and editor-native orchestration. These capabilities address different failure points in AI-assisted software engineering. Treating them as one generic “Copilot update” hides the architecture engineers actually need to evaluate.

A reproducible adoption checklist for engineering teams

1. Inventory access. Record plan, supported models, client versions and organization policies; do not assume rollout is identical for all employees.
2. Define a task matrix. Choose small bug fixes, multi-file changes and one constrained agent task; apply the same automated tests to candidate models.
3. Enforce boundaries. Test a denied filesystem path, blocked network request and restricted credential access in a safe environment before trusting a sandbox configuration.
4. Trace one complete run. Confirm that monitoring captures enough data to diagnose model and tool operations without recording confidential prompts unnecessarily.
5. Review collaboration handoffs. Create a sample issue from a controlled Slack or Teams thread; verify repository scope, duplicate suggestions and source attribution.
6. Measure the right outcomes. Compare verified changes, review minutes, failed attempts, end-to-end latency and AI Credits—not generated tokens or marketing claims alone.

A model's cost and security profile are products of the entire workflow. For deeper model comparison and alternative tool architectures, read SXF's GitHub Copilot alternatives, AI Agent Security and the agent reliability framework.

Frequently asked questions: GitHub Copilot September 2026 releases

What changed in GitHub Copilot during the September 21, 2026 week?

The roundup lists four models, Copilot app sandboxing and telemetry, Slack/Teams collaboration improvements, JetBrains controls and VS Code 1.139 remote agent sessions. The roundup itself was published on September 25.

Which new Copilot models were listed and for whom?

Claude Opus 5.5 and GPT-6 Sol were listed for Pro+, Max, Business and Enterprise. GPT-6 Luna and Grok 4.7 were listed for Pro, Pro+, Max, Business and Enterprise. These are September announcement details; confirm live support separately.

Is Copilot local sandboxing still in preview?

It was a public preview in the September 23 Copilot app announcement. GitHub announced general availability on October 7, 2026 for the Copilot app, Copilot CLI and supported VS Code Agent Host sessions.

Can Copilot use Slack messages and Teams context to create issues?

GitHub documented supported conversation context, similar-issue checks and direct links from created work to its source discussion. Permissions, integration setup and human review still matter.

Does VS Code 1.139 support remote Dev Container agent sessions?

Yes, including eligible SSH, Tunnel and WSL project environments; Docker and a supported Dev Container configuration are required on the remote host. GitHub and VS Code described gradual feature enablement.

Do sandboxing and OpenTelemetry eliminate agent security risks?

No. Sandboxing constrains tool access, while telemetry shows what happened. Neither automatically verifies the resulting code, requirements, secret handling or repository policy.

Primary sources and verification notes

This report is a source-attributed interpretation of GitHub and VS Code release materials. No fabricated performance numbers are used. Release-week facts were cross-checked against their dated announcements; availability and billing should always be rechecked against live GitHub documentation. Research review: November 8, 2026.

1. GitHub Changelog, September 25: Weekly releases — September 21 — primary roundup.
2. GitHub, September 22: Claude Opus 5.5 in Copilot — model availability and launch evidence.
3. GitHub, September 23: Local sandboxing preview — setup and scope at launch.
4. GitHub, October 7: Local sandboxing GA — subsequent status update.
5. GitHub, September 22: OpenTelemetry — traces, managed settings and content-capture defaults.
6. GitHub, September 25: Slack and Teams update — context and reliability.
7. GitHub, September 22: JetBrains 1.18.0 update — approvals, skills and agent controls.
8. Microsoft VS Code 1.139 release notes — remote Dev Containers and session management.
9. GitHub Docs: Supported AI models — current availability.
10. GitHub Docs: Models and pricing — per-model usage billing.
11. GitHub Docs: Model availability policy — organization and enterprise model controls.

Related SXF analysis: Copilot October 5 weekly updates · VS Code Copilot September releases · Opus 5.5 in Copilot · Copilot alternatives.