GitHub Copilot Gets GPT-5.4: Native Computer Control, 1M Token Context, and What Every Developer Needs to Know

What Actually Changed from GPT-5.3-Codex to GPT-5.4

GPT-5.3-Codex, released to Copilot in February 2026, was a strong code-specialist model but remained fundamentally a single-application tool — it could edit files, run terminal commands, and navigate a repository, but it could not reach outside the IDE context to interact with other applications on the developer's machine. GPT-5.4 breaks that boundary with native computer control.

The context window expansion is equally significant. GPT-5.3-Codex operated within a 128K token limit — enough for large files and multi-file sessions but insufficient for monorepos or deep cross-cutting analysis. GPT-5.4's 1 million token context means the entire source tree of a typical mid-sized production application fits within a single session. This eliminates the retrieval-augmented generation (RAG) step that many Copilot extensions required to work around context limits.

GPT-5.3-Codex simultaneously received Long-Term Support (LTS) designation on March 18, 2026 — signalling that enterprise teams on regulated deployment cycles can pin to it for 18–24 months while evaluating GPT-5.4 for production use.

Native Computer Control: How It Works

GPT-5.4's computer control works through a screenshot-observe-act loop. The model takes a screenshot of the current desktop state, interprets the visual layout to identify interactive elements (buttons, input fields, menus), issues simulated mouse click or keyboard input events, and re-evaluates the updated screenshot. This loop repeats until the task goal is satisfied or an error condition triggers a pause for human review.

In practice, this allows Copilot agent sessions to span applications that have no API or CLI interface. A developer can instruct Copilot to: open a Figma design file, extract colour values and component names, create matching TypeScript constants in the IDE, run a test suite, read the test output in the terminal, and file a GitHub issue — all in a single agent session without switching focus or writing glue scripts.

// Copilot agent prompt (VS Code chat):
"Pull the latest design tokens from Figma (the 'Design System v3' file),
update /src/tokens/colors.ts and /src/tokens/spacing.ts to match,
run `npm run test:tokens` and fix any failures, then open a PR."

// GPT-5.4 execution trace:
[1] Takes screenshot → identifies Figma icon in taskbar
[2] Clicks Figma → navigates to 'Design System v3'
[3] Screenshots token panel → parses colour/spacing values visually
[4] Returns to VS Code → edits colors.ts and spacing.ts
[5] Opens integrated terminal → runs `npm run test:tokens`
[6] Reads test output → detects 2 failures → fixes type mismatches
[7] Runs tests again → all pass
[8] Calls `gh pr create` via terminal → PR opened

Security note: Computer control runs in a sandboxed session scoped to the IDE workspace. It cannot access files or applications outside the defined workspace boundary without explicit user approval at each step. GitHub's governance settings for Business and Enterprise allow administrators to restrict computer control to approved application lists.

Thinking vs Pro: Which Variant to Use

GPT-5.4 ships in two variants within Copilot. GPT-5.4 Thinking adds an explicit chain-of-thought reasoning phase before generating output — the model reasons through the problem step-by-step before committing to an implementation. GPT-5.4 Pro skips the reasoning phase and optimises for throughput and cost in production workloads.

The practical heuristic: use Thinking when the problem is ambiguous or high-stakes (you'd review the output carefully regardless), and Pro when you need volume (running agent sessions in CI, generating tests across a codebase, scaffolding new services).

Enabling GPT-5.4 in Your IDE

GPT-5.4 requires minimum IDE extension versions. Update your Copilot extension first — older versions fall back to GPT-5.3-Codex silently.

// In VS Code Copilot Chat panel — click model selector dropdown:
// Options: GPT-5.4 Thinking | GPT-5.4 Pro | GPT-5.4 mini | GPT-5.3-Codex (LTS)

// Or set default in settings.json:
{
  "github.copilot.chat.defaultModel": "gpt-5.4-pro",
  // options: "gpt-5.4-thinking" | "gpt-5.4-pro" | "gpt-5.4-mini" | "gpt-5.3-codex"
  "github.copilot.agent.computerControl": true,
  "github.copilot.agent.contextWindow": 1000000
}

GPT-5.4 mini: The Speed Layer for Codebase Exploration

GPT-5.4 mini, GA'd March 17, is the model you should use for real-time inline completions and codebase navigation. It delivers the lowest time-to-first-token of any Copilot model and is specifically optimised for grep-style tool use — the read-only file and symbol searches that dominate interactive coding sessions.

In practice this means: use mini for the chat sidebar and inline suggestions during active coding, and switch to GPT-5.4 Thinking or Pro only when invoking a longer agent session. Many developers will find a mixed-model workflow optimal: mini for the 80% of routine interactions, Thinking for the 20% that require deep reasoning.

{
  // Fast completions and inline suggestions
  "github.copilot.inlineSuggest.model": "gpt-5.4-mini",

  // Sidebar chat — default to Pro, switch manually to Thinking
  "github.copilot.chat.defaultModel": "gpt-5.4-pro",

  // Agent mode (longer autonomous sessions)
  "github.copilot.agent.model": "gpt-5.4-thinking",

  // Computer control for cross-app workflows
  "github.copilot.agent.computerControl": true
}