For the past year, software developers working inside Visual Studio Code have faced a growing paradox of choice. Every few weeks, a new frontier or open-weight coding model lands in the model picker—each with different strengths across fast autocomplete, architectural planning, unit-test generation, and security code review. While VS Code 1.143 introduced a side-by-side “Agent Bake-Off” to let developers manually test prompts across up to four models at once, manually refereeing models on every ticket quickly becomes a tax on developer focus.
With the release of Visual Studio Code 1.144 alongside GitHub Universe on October 28, 2026, Microsoft is moving from manual model comparison to automated multi-model orchestration. The headline feature is the general availability of GitHub Copilot Fusion , which combines Project HydraFusion’s multi-model routing engine with task-aware execution patterns. Instead of forcing a developer to pick a single model for an entire session, Copilot Fusion dynamically selects one or several models—using separate models to generate, review, and critique code—based on the complexity of the task and an adjustable effort setting (Standard, High, or Max).
Beyond multi-model code orchestration, VS Code 1.144 expands the boundary of what an IDE agent can touch. The update introduces experimental Computer Use (toggled via /computer on and /computer off), allowing locally running Copilot harness sessions to interact directly with desktop applications on Windows and macOS, alongside native GPT Image 2.5 Flare image generation inside Copilot Chat, event-driven cloud automations, and version 1.0 of the open Agent Host Protocol .
For individual developers and enterprise engineering managers, VS Code 1.144 turns the editor into a full-stack agent workstation—while introducing new billing and desktop-security rules that teams need to configure before enabling Max effort or /computer on.
Copilot Fusion General Availability: Built on Project HydraFusion, Copilot Fusion automatically routes tasks across one or multiple models (generating, reviewing, and critiquing solutions) across Standard, High, and Max effort tiers (High and Max require Copilot Pro+, Max, Business, or Enterprise).Per-Model Billing & 30% Launch Discount: Users are billed at the standard token/request rate for each model that Copilot Fusion invokes during a multi-model pass; paid Copilot plans receive a 30% discount on Fusion usage through November 30, 2026 .Experimental Desktop Computer Use (/computer on): Local Copilot harness sessions on Windows and macOS can now interact directly with desktop GUI applications after explicit user authorization, controlled via /computer on and /computer off chat commands.In-Chat Image Generation (GPT Image 2.5 Flare): Developers can generate and edit website graphics, icons, backgrounds, and UI wireframes directly inside Copilot Chat using GPT Image 2.5 Flare (consuming additional AI credits, with an optional visible AI watermark setting).Cloud Automations & Agent Host Protocol 1.0: Preview cloud automations trigger background agents on repository events (such as issue assignments or pull-request activity), while Agent Host Protocol v1.0 standardizes communication between local and remote agent sessions. Inside Copilot Fusion: Multi-Model Generation, Review, and Critique
Why use three models on a single coding task when one frontier model can write syntax on its own? Because different model families fail in different ways. A high-speed model excels at scaffolding boilerplate routes or updating TypeScript interfaces in two seconds, but frequently misses subtle edge cases in state management or SQL concurrency. Conversely, invoking a heavyweight reasoning model for every one-line CSS tweak wastes both wall-clock time and premium API credits.
Copilot Fusion automates that division of labor. When a developer submits a prompt at Standard effort, Fusion routes simple edits to a single fast model. When the developer escalates a complex architectural refactor or concurrency bug to High or Max effort, Fusion orchestrates a multi-stage pipeline: one model drafts the implementation, a second independent model reviews the diff for logic and security flaws, and a critique pass synthesizes the final verified patch before presenting it in the editor.
VS Code 1.144 Feature
Execution Mode
Billing & Plan Requirement
Best Practical Use Case
Copilot Fusion (Standard) Single task-matched model selection
Standard single-model request rate (30% off thru Nov 30)
Everyday feature scaffolding, unit tests, and quick refactors
Copilot Fusion (High / Max) Multi-model generate -> review -> critique loop
Billed per model invoked (Pro+, Max, Business, Enterprise)
Complex multi-file architecture changes, auth, and concurrency fixes
Agent Bake-Off (Expanded) Parallel side-by-side prompt comparison
Billed per selected model (now works without an open repo/Git commit)
Benchmarking how new models handle a tricky prompt before committing
Desktop Computer Use (/computer on) Local OS GUI control (Windows & macOS)
Local Copilot harness + explicit user authorization prompt
Automating end-to-end desktop app testing, Electron UIs, and native dialogs
Copilot Fusion (Standard)
Execution Mode Single task-matched model selection
Billing & Plan Requirement Standard single-model request rate (30% off thru Nov 30)
Best Practical Use Case Everyday feature scaffolding, unit tests, and quick refactors
Copilot Fusion (High / Max)
Execution Mode Multi-model generate -> review -> critique loop
Billing & Plan Requirement Billed per model invoked (Pro+, Max, Business, Enterprise)
Best Practical Use Case Complex multi-file architecture changes, auth, and concurrency fixes
Agent Bake-Off (Expanded)
Execution Mode Parallel side-by-side prompt comparison
Billing & Plan Requirement Billed per selected model (now works without an open repo/Git commit)
Best Practical Use Case Benchmarking how new models handle a tricky prompt before committing
Desktop Computer Use (/computer on)
Execution Mode Local OS GUI control (Windows & macOS)
Billing & Plan Requirement Local Copilot harness + explicit user authorization prompt
Best Practical Use Case Automating end-to-end desktop app testing, Electron UIs, and native dialogs
Notice that VS Code 1.144 also removes a frustrating limitation from the side-by-side Agent Bake-Off introduced in 1.143: developers can now run multi-model comparisons without needing an open project folder or an existing Git commit (though Dev Containers remain unsupported for side-by-side comparisons).
Desktop /computer on and GPT Image 2.5 Flare: Full-Loop UI Development
Two features in VS Code 1.144 specifically target the friction frontend and full-stack developers experience when switching between their code editor, design tools, and native desktop windows:
Inline Asset Creation with GPT Image 2.5 Flare: Previously, if a developer building a landing page or mobile app needed a placeholder hero illustration, SVG-style icon set, or wireframe mockup, they had to leave VS Code, generate assets in an external tool, download the files, and import them into the repository. With GPT Image 2.5 Flare integrated into Copilot Chat, developers can generate or modify existing images in-place. Because image generation consumes extra AI credits beyond text chat, teams should reserve it for rapid UI prototyping and enable the optional visible AI watermark setting on staging builds so placeholder graphics are never accidentally shipped to production as final brand art.Experimental Desktop Computer Use (/computer on): Browser-only preview tools work fine for web apps, but fail when testing native macOS/Windows apps, Electron desktop shells, system tray utilities, or multi-window OAuth popups. By typing /computer on in a local Copilot harness session, developers authorize the agent to inspect and interact with desktop applications outside the VS Code window—and immediately revoke that access with /computer off once the test completes.
Practical Cost and Security Governance for Engineering Teams
While Copilot Fusion and /computer on significantly expand developer velocity, they also introduce two operational risks if rolled out without guardrails: multiplied token consumption and unintended desktop GUI actions .
Because Copilot Fusion charges for each underlying model it invokes during a High or Max orchestration pass, a developer who leaves Fusion locked on Max effort for routine coding can burn through monthly premium request allowances three times faster than expected—especially once Microsoft’s temporary 30% promotional discount expires after November 30, 2026.
The 4-Step Enterprise Implementation Playbook
Default to Standard Effort and Reserve High/Max for PR Pre-Flight: Train developers to keep Copilot Fusion set to Standard during iterative coding and switch to High or Max specifically when designing multi-file architecture or running a pre-PR multi-model critique pass. Take advantage of the 30% discount window through November 30 to benchmark exactly how many credits Max consumes on your codebase.Enforce Strict Session Hygiene Around /computer on: Treat the /computer on command like sudo. Developers should close password managers, production cloud consoles, and internal Slack/email windows before granting a local Copilot harness desktop GUI control, and immediately execute /computer off as soon as the desktop UI verification finishes.Turn On AI Watermarks for GPT Image 2.5 Flare Prototypes: Enable the visible AI-generated watermark setting in workspace preferences whenever developers use GPT Image 2.5 Flare to generate icons, backgrounds, or wireframes. This ensures design and legal reviewers can instantly spot temporary AI placeholder assets before a release goes live.Move Repetitive Triage to Event-Driven Cloud Automations: Instead of burning local IDE time on routine issue reproduction or changelog updates, configure VS Code 1.144’s preview Cloud Automations in the Agents window to trigger background GitHub-hosted agent sessions automatically when issues are assigned or pull requests are opened.
Next Steps
