Why does Codex become noticeably slower when using GPT-5.6 Sol?

[Heyu2002]

When using GPT-5.6 Sol Ultra, I find the response speed significantly slower than with GPT-5.5, sometimes even very simple project tasks consume an excessive amount of time.
What puzzles me is that this does not reflect clearly in token consumption—something I find confusing. In my experience, I've even noticed that for extremely simple questions, GPT-5.6 attempts to create multiple subagents to assist, perhaps to save tokens? However, in such cases, many of these subagents frequently fail to connect and require reconnection, which seems highly abnormal.

[Mahkhmood9]

Sol Ultra , should be used rarely

use your allowance wisely

I recomend using Terra-medium-thinking , but I recomend start with the lowest settings
gpt5.5 is really over kill --- gpt 5.6 has seen all of the use cases where 5.5 fails --- so you wont be disapointed from the smaller models

dont get tempted with the highest settings.....

but look you are paying for the sub, if you want you can use , sol and the highest settings

all I am saying you are going to have a good time using other models too

i do not recomend building with sol ,, I think its more wisely using it only for code review/ security review/ compacting and refactouring

i recomend you chekc https://learn.chatgpt.com/use-cases/refactor-your-codebase
for refectoring your codebase and making it a little bit mroe lean

as per speed, there are better optimied models for it

nvidia is expected to release their new NVIDIA GROQ LPX , which I personally expect them to release it with support for the NVFP4 , without , with the power of what GROQ and their SRAM tricks - without the errors and issues - we seen when not running nvidia hardware for infrence, despite TPU and 600 people working on compilers for their hardware... despite results in so called ai infrence chips... with nvidia products debicated for servers , I think whats beautiful is that they can built a cpu, and also remove a lot of support for a lot of things... something that AMD cant afford to for example,, and its not like RISCV is truly open source - you still gonna pay for someone else i.p

as per TPU, its a good product, google can generally afford a lot of infrence, almost any google search will be ingensted again a large llm model

[Mahkhmood9]

as per multiple subagents , I am not a fan of it either -- I rather use a small model -- but perhaps it can hint on good future ideas

each "agent" mean a user, which going to harm infrence for users... as/or -not SOL === 20users-who-use-TerraORluna

I dont see the reason for using SOL, as you can be just as happy with Terra or Luna - try use the next 9 prompts with it

for example I am able to very easily navigate the chromium code base with it, which can be complex... I assume Luna can work great with smaller codebases -

[ethan-2937]

Sincerely asking: which one would be better? Terra medium or luna high/xhigh? I tried sol. For me, it seems better than 5.5, but it's too slow ):

[alexkkork]

good work takes time

[Mahkhmood9]

thats not true, Terra is blazing fast, , for me even better than SOL. honestly just ignore SOL even exsit