GitHub Copilot: Unified Inline Suggestions Model in VS Code
GitHub CopilotGitHub Copilot in VS Code now runs a single 3-in-1 unified model for inline suggestions, replacing the separate models that previously handled completion-style ghost text, near-cursor next edit suggestions, and longer-range edits elsewhere in the file. The unified model shows 13% less ghost text overall while cutting suggestion dismissals by 10.1%, since one model can weigh all three suggestion types together instead of chaining independent requests. It powers a smoother Tab-Tab-Tab workflow that propagates an edit pattern to other relevant spots in a file, and it is available now for paid Copilot users running the latest VS Code with next edit suggestions enabled.
Key Takeaways
- GitHub Copilot replaced three separate models for ghost text, near-cursor edits, and long-distance edits with a single unified "3-in-1" model in VS Code.
- Despite showing 13% less ghost text, the new model cut suggestion dismissals by 10.1%, meaning fewer but more relevant suggestions.
- The change is the second phase of a two-part unification effort GitHub Copilot's engineering team documented publicly across September 16 and September 23, 2026.
- Training the 3-in-1 model took just 35 RL runs and 10 A/B flight candidates, far fewer than the 170-plus runs the earlier 2-in-1 model required.
- The Tab-Tab-Tab workflow, which propagates an accepted edit to other relevant spots in a file, benefits directly from the unified reasoning.
- The feature is live now for paid Copilot subscribers on the latest VS Code with next edit suggestions enabled, no separate opt-in required.
Sources & Mentions
2 external resources covering this update
A Single Model for Every Kind of Suggestion
GitHub Copilot's inline suggestions in VS Code used to come from three separate systems working side by side: a completions model that produced ghost text as a developer typed, a next edit suggestions (NES) model that proposed changes near the cursor, and a third model that handled edits farther away in the same file. Each system ran its own request, and VS Code had to orchestrate the hand-offs between them. With the release of VS Code 1.139, GitHub Copilot consolidates all three into one unified model, referred to internally as the "3-in-1" model, that considers every kind of suggestion in a single pass.
This is the second phase of a broader unification effort GitHub Copilot's team described across a two-part blog series published September 16 and September 23, 2026. The first phase merged next edit suggestions and long-distance edits into a "2-in-1" model; the second phase folded the completions model into that same system.
Why Fewer Suggestions Beat More
Counterintuitively, the unified model shows developers 13% less ghost text overall than the previous multi-model setup, yet suggestion dismissals dropped by 10.1% with no regressions in Copilot's other quality metrics. GitHub's engineering team attributes this to the unified model being able to "optimize for the best edit that fits the moment" rather than being boxed in by the artificial boundaries between separate specialist models. When one system can reason about ghost text, near-cursor edits, and distant edits together, it can choose not to show a suggestion at all when a different one would serve the developer better, instead of firing off a suggestion from whichever specialist model happened to trigger first.
The unification also simplified the underlying training process. GitHub Copilot's team reports the 3-in-1 model needed only 35 reinforcement learning runs and 10 A/B flight candidates to reach production quality, compared to more than 170 RL runs and 15-plus supervised fine-tuning runs for the earlier 2-in-1 model, a sign that lessons from unifying two systems compounded quickly when unifying a third.
What Developers Will Notice
In practice, the change should make Copilot's suggestions in VS Code feel more continuous. Completion-style ghost text now appears more progressively as a developer types, proactive ghost text can appear after a cursor move alone if Copilot has enough context to infer the next intended change, and the existing Tab-Tab-Tab workflow, where accepting one suggestion carries the same edit pattern to other relevant locations in the file, benefits from the model reasoning about all three suggestion types at once rather than stitching together separate outputs.
Availability
The unified inline suggestions experience is live now for paid GitHub Copilot users in VS Code. GitHub Copilot's team recommends running the latest version of VS Code and making sure next edit suggestions are turned on to get the full benefit of the unified model.