AI Coding News: July 31, 2026 — Copilot Retires Gemini 2.5 Pro and 3 Flash

July 31, 2026 closes the month with a clean model cutover rather than a splashy new product. GitHub Copilot officially retires Gemini 2.5 Pro and Gemini 3 Flash today, capping a month-long deprecation notice. Alongside that, GitHub tightened up automation and governance: Copilot CLI no longer needs a personal access token inside GitHub Actions, Business and Enterprise admins get a default-on policy for new models, and remote-controlled CLI sessions can now be locked to managed devices. Outside GitHub's orbit, Visual Studio picked up a new SDK-based Copilot agent, Cognition shipped self-hosted "Outposts" for Devin, and Google pushed a stability release for Gemini CLI. Here's what changed.

GitHub Copilot Retires Gemini 2.5 Pro and Gemini 3 Flash

As flagged in GitHub's July 2 deprecation notice, Gemini 2.5 Pro and Gemini 3 Flash are pulled from every Copilot surface today — Copilot Chat, inline edits, Ask and Agent modes, and code completions all lose access to both models simultaneously. Anything still pinned to either one needs to move to Gemini 3.1 Pro or Gemini 3.5 Flash, the models GitHub is steering customers toward; Google's newer Gemini 3.6 Flash, which landed in Copilot's model picker on July 21, is also available as a replacement. Workflows, custom agents, or CI jobs that hardcode a model name will start failing today rather than degrading gracefully, so it's worth checking any Copilot configuration for the old model IDs before end of day.

Copilot CLI Drops Personal Access Tokens Inside GitHub Actions

Copilot CLI can now authenticate inside GitHub Actions using the workflow's built-in GITHUB_TOKEN instead of a stored personal access token, closing a gap GitHub started addressing with agentic workflows back in June. Enabling it just requires the "Allow use of Copilot CLI billed to the organization" policy plus a copilot-requests: write permission on the workflow — no PAT to mint, rotate, or leak. AI credits consumed this way bill straight to the organization, which also gives finance teams a cleaner line on CLI-driven CI spend.

Enterprise Governance: Default Model Enablement and Locked-Down Remote Control

Two admin-facing changes rounded out GitHub's week. A new default-availability policy means generally-available Copilot models switch on automatically for Business and Enterprise orgs instead of waiting on manual approval — open-weight models like Kimi K2.7 and models outside GitHub's data retention agreement, such as Fable 5, are excluded regardless of the setting. The policy is live now but has no effect for 28 days, giving admins who want manual approval until August 26 to switch it to disabled. Separately, enterprises can now restrict which devices are allowed to host remotely controlled Copilot CLI sessions, with a remoteControl managed setting that supports requireSSO, disabled, or unrestricted modes — useful for teams that opened up mobile and web remote control back in May but want it locked to known hardware now.

GitHub Copilot in Visual Studio Gets an SDK-Based Agent

Visual Studio's July update introduces a new Agent (Preview) built directly on the GitHub Copilot SDK, aimed at getting more tasks right on the first pass with less back-and-forth and shorter, easier-to-scan responses. It ships alongside built-in skills authored by Microsoft's own .NET and Azure teams, available from the "Built-in" category of the tool picker — a way of baking framework-specific expertise into the agent rather than relying entirely on general model knowledge.

Devin Ships Outposts for Self-Hosted Runs

Cognition introduced Devin Outposts on July 21, letting Devin Cloud sessions execute on infrastructure you control — a Mac mini, a GPU box, a VM on a private network, or a Kubernetes cluster — while planning and inference stay in Cognition's cloud. Outposts are created in Devin Cloud and connected through an Outpost worker run via the Devin CLI, with orchestration support for platforms like Modal, E2B, Daytona, and NVIDIA Brev. It's aimed at teams that need Devin working against internal services, private data, or specialized hardware it couldn't otherwise reach; the feature ships disabled by default and requires contacting Cognition to turn on.

Gemini CLI v0.50.0 Ships Stability Fixes

Google's Gemini CLI moved to v0.50.0 with a smaller, maintenance-focused release: fixes for JSON and IPYNB handling in write_file, more reliable previous-intent labels, crash hardening around PTY resize events, and a switch to the Flash GA model when the relevant experiment flag is set. It's a quieter release than July's earlier Antigravity consolidation, but it keeps the CLI's core file-editing path more predictable for scripted and agentic use.

The Takeaway

Nothing shipped today rewrites what these tools can do — it's cleanup. GitHub is retiring models on schedule while handing admins both more automatic and more locked-down control at the same time, Visual Studio and Devin are both investing in agents that lean on curated, first-party expertise rather than raw model capability alone, and Gemini CLI's release is pure stability work. If you have anything — a workflow, a script, a saved prompt — pointing at Gemini 2.5 Pro or Gemini 3 Flash inside Copilot, that's the one action item with a hard deadline: today.

References & Further Reading

• Upcoming deprecation of Gemini 2.5 Pro and Gemini 3 Flash — GitHub Changelog
• Gemini 3.6 Flash is now available in GitHub Copilot — GitHub Changelog
• Copilot CLI no longer needs a personal access token in GitHub Actions — GitHub Changelog
• Default model enablement for Copilot Business and Enterprise — GitHub Changelog
• Limit remote control to managed devices — GitHub Changelog
• GitHub Copilot in Visual Studio — July update — GitHub Changelog
• Introducing Devin Outposts — Cognition
• Gemini CLI Release Notes — Google