AI updates are happening so fast it's hard to keep up. That's a huge missed opportunity.
Codex changed multiple times in July 2026 alone. If your Codex is still on default settings, you aren't even seeing half of what GPT-5.6 Sol can do.
Codex can change its behavior with short commands and settings. That hasn't changed.
However, the internal mechanics were completely overhauled in July.
How safety checks are performed.
How connection destinations are selected.
Internal browser functionality.
Operation from smartphones.
And the model itself.
Knowing these makes the same Codex feel like a completely different tool. I recommend bookmarking these 10 features and trying them one by one.
Before we dive in, I am currently giving away 55 major bonuses, including a Codex textbook, setup guides, and monetization methods for free. You can get them immediately from the link below if you haven't already.
↓
https://utage-system.com/line/open/cwgwX1a35XDK?mtid=FNAamIuYaEet
Now, let's get to the main topic.
1. Official OpenAI Plugin for Running Codex via Claude Code
Tools from rivals OpenAI and Anthropic have officially joined forces in the user's hands.
OpenAI released an official plugin for Claude Code on GitHub called codex-plugin-cc.
Installation takes just two lines:
1/plugin marketplace add openai/codex-plugin-cc2/plugin install codex@openai-codex
By typing /codex:review, you can run a Codex review from within Claude Code. There are also /codex:rescue and /codex:transfer commands.
You can offload tasks that get stuck in Claude Code to Codex. Plan in Claude Code and implement in GPT-5.6 Sol. If one gets stuck, throw it to the other. This is now a one-command process via the official plugin.
Since you can make the two strongest models collaborate, even those who usually only use Claude Code should try this plugin.
2. Correct Usage of /goal
Prompts passed to /goal are set directly as the objective. If you throw a vague sentence, it will run wild with that vagueness.
First, have Codex research the task to clarify it, then pass the objective via /goal. This alone reduces deviations once it starts running.
1Please write a /goal prompt that allows this task to be executed comprehensively and completed. Then, share that goal prompt with me.
Sending the above prompt will generate a prompt for /goal, which is very convenient.
3. Batch Cookie Import for Internal Browser
X, Reddit, LinkedIn. Every time you had Codex research something, you had to reopen the browser and log in again.
This hassle disappeared with the July 9th update.
Codex's internal browser can now batch import cookies and passwords from Chrome.
Once imported, you can operate X, Reddit, LinkedIn, etc., within Codex while remaining logged in.
This is different from the "Codex Chrome Extension" released in May. That one stayed in the Chrome profile; this time, the app's internal browser itself has been strengthened. For research tasks, this change alone significantly improves perceived speed.
4. Enhanced Dangerous Command Detection
On July 16, detection for dangerous commands was strengthened. It now catches forced deletion commands like rm -rf in a wider variety of patterns and clearly displays the reason for rejection.
It's a subtle but effective change. Codex itself will stop accidental executions while you aren't looking.
5. Default Activation of Remote Plugins
It's common to install plugins and then forget they exist.
On July 8, this was addressed with an automatic remote plugin suggestion feature. Codex now suggests relevant plugins based on the conversation and can even handle the installation.
/plugins has also been organized into OpenAI Curated, Workspace, and Shared with me. The problem where having more plugins made them harder to use has been resolved.
6. Automation of MCP Tool Search
It's tedious to manually choose which MCP to use when you have many connected. This was also solved on July 8.
MCP is the standard for connecting external tools to Codex. Tool searching is now on by default. Codex automatically searches for and loads only the necessary tools from your connected library every turn.
You no longer need to choose which tool to use; Codex will select the appropriate MCP tool automatically.
7. In-Chat Visualization Feature
For those who thought complex information would always result in long walls of text, there is good news.
A visualization feature has been added to Codex chats. You can check tables, diagrams, and interactive elements directly within the task without moving to another screen. Results can now be visualized right on the smartphone task screen.
8. Differentiating Reasoning Effort
The levels of thinking depth have increased: low, medium, high, xhigh, and the new "max." Max is for difficult problems where a single model needs to think as deeply as possible.
A common misunderstanding is "Ultra." Ultra isn't a depth level; it's a completely different system that runs multiple sub-agents in parallel for division of labor.
Use High as a base, and switch to xhigh or max only when necessary. Ultra should be reserved for when multi-agent effects are expected, as it consumes many tokens.
9. Choosing Between Sol, Terra, and Luna
The previous points were peripheral; this is the core. GPT-5.6 isn't one model; it's divided into three tiers.
Sol is the flagship for complex, high-quality work. Terra is the balanced model, offering GPT-5.5 level performance at a lower cost. Luna is the fastest and cheapest, meant for execution.
The practical workflow spreading on X is: let Sol handle planning and review. To save tokens, let Terra handle sub-agent execution, and let Luna handle simple tasks like file checking. Luna is reported to be 1.3x faster and 2.5x cheaper than Terra, making it harder to hit usage limits.
10. Setting Default Response Styles in config.toml
If you find Codex's replies too cold or too long, the place to fix it is the config file, not the prompt.
Just add these two lines to ~/.codex/config.toml:
personality = "pragmatic"
model_verbosity = "low"
"Personality" sets the style (friendly or pragmatic). Pragmatic makes responses business-like and direct. "Model_verbosity" sets length (low, medium, high). Low cuts out unnecessary introductions. It's much faster than adding "be concise" to every prompt.
Conclusion
These are the 10 divine features of Codex. You don't have to learn them all at once. Bookmark this article and try them one by one to transform your Codex experience.
I regularly post about the latest AI info and monetization methods. If this was helpful, please follow or quote. Also, don't forget to grab the 55 bonuses from the link below!
↓
https://utage-system.com/line/open/cwgwX1a35XDK?mtid=FNAamIuYaEet
Also, for 3 days only (July 17-19), I'm holding a free AI x High-Ticket Affiliate Automation Seminar. I'll explain how to monetize quickly from zero, so book your spot before they're gone!
→https://utage-system.com/page/EuRvqtG80p2R?ftid=P7rCma4bLq8e
