Expanding the Codex Experience in VSCode: A Thorough Review of GPT-5's Thinking Mode

In my previous article, I introduced how to install “Codex – OpenAI’s coding agent” in VS Code to use GPT-5, as well as my experience switching between “Thinking Mode” and “Approval Mode.”
In this follow-up, I’ll take a deeper dive into what Thinking Mode actually entails, how Approval Mode works, and tips for using it with cloud integration.

---

Feature and specification statements were checked on July 31, 2026 against Introducing GPT-5. The article distinguishes what the official source confirms from editorial organization and recommendations.

Basic Design of the Codex Extension

• Install it from the VSCode Marketplace.
• It’s ready to use simply by signing in with your OpenAI account (no need to manually enter an API key).
• In addition to pair work within the editor, it features a mechanism to delegate tasks to the cloud.

This two-tiered structure enables seamless transitions from lightweight code completion to large-scale refactoring.

---

A Fresh Look at GPT-5’s Thinking Modes

From a practical perspective, the differences between “Instant,” “Balanced,” and “Thinking” modes—which you can experience in Codex—can be broken down as follows:

• Instant (Immediate)

  • Responds quickly with minimal search
  • Ideal for code completion and short explanations
  • Can be used as if “asking a colleague a quick question”

• Balanced (Main Mode)

  • Strikes a balance between speed and accuracy
  • Handles code reviews and article drafts well
  • Feels like a “reliable pair programmer”

• Thinking (In-Depth)

  • Takes a few seconds to respond but offers rich logical reasoning
  • Strong for design reviews and bug investigations
  • Feels like “consulting a senior engineer at length”

The key here is that you can choose the trade-off between “speed” and “depth” yourself.
Switching modes based on the nature of the task is what drives productivity.

---

Approval Mode for Safety

Codex also features tiered permission settings.

• Chat: Conversation only. Does not make unauthorized changes to code or the environment.
• Agent: Automatically makes changes within the working directory, but external access requires approval.
• Agent (Full Access): Can manipulate the network and a wide range of areas. Recommended for use in experimental environments.

This helps alleviate concerns such as, “I can’t trust AI because I’m afraid of making mistakes.”

---

What You Can Do with Cloud Integration

Codex integrates not only with local environments but also with cloud execution environments.

• Execute large-scale test generation and refactoring externally
• Review the results before applying them locally
• Parallelize tasks in the background without interrupting your train of thought

This mechanism reinforces the division of labor where “humans set the direction, and AI handles the work.”

---

Practical Usage Patterns

1. Specification Organization → Thinking Explicitly state assumptions and constraints to reduce oversights.
2. Initial Implementation → Balanced Ensure the shortest working code and a minimum set of tests.
3. Rush of Minor Tweaks → Instant Process naming and comment revisions in seconds.
4. Bug Tracking → Thinking Identify side effects of changes and test cases.
5. Delegation to the Cloud → Balanced/Thinking
   Outsource heavy tasks and absorb only the results.

---

Summary: AI “Gear Shifting” at Your Fingertips

What I experienced through the Codex extension was a new workflow where you can switch GPT-5 between “speed-first” and “depth-first” modes.

• Light tasks: Instant
• Main workflow: Balanced
• Design and research: Thinking

Furthermore, by combining approval mode and cloud integration, you can safely and efficiently utilize AI as a “partner.”

👉 Related Articles:

• Practical Examples of Article Writing with ChatGPT
• Connecting AI and Development

---