Hands-on with Codex in VS Code: A Practical Look at GPT-5 Thinking Modes

Codex Deep Dive

In our earlier walkthrough we set up Codex inside VS Code. This time we take a closer look at how GPT-5's thinking modes behave in daily development, how approval levels keep projects safe, and when to hand work to the cloud runtime.

Quick Recap: Codex in VS Code

Install the extension from the VS Code Marketplace.
Sign in with your ChatGPT (OpenAI) account; no manual API keys.
Run tasks locally inside your workspace or delegate bigger jobs to Codex's managed cloud sandbox.

That split lets you move from quick edits to large refactors without context switching.

How the GPT-5 Modes Feel

Instant: Delivers answers in a heartbeat. Perfect for short completions, quick translations, and tiny refactors. It feels like leaning over to ask a colleague a simple question.
Balanced: Blends speed with reasoning. Use it for code reviews, writing docs, or assembling mid-sized features. Think of it as your daily pair programmer.
Thinking: Slower on purpose. It traces dependencies, weighs trade-offs, and drafts structured plans. Reach for it when debugging snarled logic or designing new architecture.

Switching modes consciously is like shifting gears in a manual car; match the mode to the road ahead and you stay productive.

Stay in Control with Approval Levels

Codex ships with three permission tiers:

Chat: Conversation only. Zero code edits or shell commands.
Agent: Can edit files and run commands in the workspace. Any attempt to touch external paths or the network triggers an approval prompt.
Agent (Full Access): Removes most guardrails. Reserve it for sandboxed projects or one-off chores where automation must keep running unattended.

Because every elevation requires consent, you always know what Codex is allowed to do.

Why Cloud Offloading Matters

Local runs are great for quick turns, but heavy tasks benefit from the managed cloud runtime:

Generate or execute large test suites without tying up your machine.
Let Codex iterate on multi-file refactors while you keep coding.
Review logs, diffs, and test results in VS Code before applying them locally.

You set intent, Codex handles the grind, and you still approve every change.

Sample Workflow

Clarify requirements with Thinking to surface assumptions and edge cases.
Draft the implementation with Balanced for a first pass and baseline tests.
Polish copy and naming with Instant for fast micro-edits.
Investigate stubborn bugs with Thinking to trace the root cause.
Offload long builds or migrations to the cloud so you are free to keep working.

Takeaway

GPT-5 inside Codex is more than autocomplete. Treat Instant, Balanced, and Thinking as gears you can shift among, pair them with the right approval level, and scale up to the cloud when needed. The result is an AI teammate that feels fast, safe, and genuinely useful.