Since GPT-5.6 was released, I have been doing extensive testing. I used it to rebuild two medium-sized agent projects from scratch and compared the results with my previous workflow. Here are some of my observations.
Pros
In terms of performance, GPT-5.6 is much better than GPT-5.5. The overall quality of outputs has improved significantly.
However, the token consumption is also much higher. In my experience, it uses around 4–5x more tokens compared with my previous workflow. Even as a 20x Pro user, after only 1–2 days of heavy usage, my quota was completely exhausted (Ultra mode)，it never happened when I was using 5.5 with extrahigh mode.
The improvement in frontend development is especially impressive. GPT-5.6 can now complete many UI and frontend changes that require almost no additional modification. This is probably the biggest improvement I noticed.
On the backend side, it was also able to solve several problems that I previously could not fix even with GPT-5.5 Extra High. Overall, this feels like a very strong upgrade in terms of capability.
Cons
The biggest issue is that it has become much slower, especially in Ultra mode.
Even with Fast mode enabled, task execution can take an extremely long time. Many tasks are measured in hours rather than minutes.
The cost is also significantly higher. Beyond the pricing issue, the current UI/UX creates serious problems.
Previously, when I used GPT-5.5 Extra with the manually orchestrated workflow, I could see my sub-agents inside the interface. I could click into each sub-agent and monitor their execution status. This was very useful for understanding whether a task was still running or already completed.
However, this workflow is not available in Ultra mode. I have no clear way to know whether the task has finished. The main agent can continue running for a very long time without providing any indication of when it will complete.
The most serious issue is resource consumption. Ultra mode frequently consumes almost all of my CPU resources. In roughly 2 out of 4 requests, my computer becomes completely frozen, forcing me to restart.
I am using a MacBook Pro M4 Pro with 24GB MBA. When only running Codex, it should not consume this level of system resources.
My Recommendation
At the moment, I do not recommend using Ultra mode.
I don’t think it is a mature workflow yet. There are still major issues with both execution reliability and user experience, not to mention the extremely high token consumption.
The underlying model capability is impressive, but the current Ultra mode feels more like a powerful experimental feature rather than a production-ready workflow. I hope future updates can improve execution transparency, resource management, and overall usability.