THANK YOU 5.6 UPDATE! Finally follows long form instruction and front door post-compaction routes.

[eah3699]

Just wanted to say, i've only been using it for a couple hours and I was worried the Codex mode was still going to bulldoze instructions and read paths from the efficiency methods and parallel reads, but it seems to follow instructions well. Been running some rigorous tests across all the models (only finishing up Sol right now) but so far so good. And High performed very well. Started off to a rocky start with xhigh burning through 41% of my usage window when i asked it the first question in the chat about the difference of token window usage from 5.5 Codex to 5.6 ChatGPTWork. Did not expect that haha. I'll keep testing but wanted to say thanks to your team.

[Mahkhmood9]

I think 5.5 was very experimental , considering they only released a single version

I also really like the way it format words, almost to the same level of claude models (I will repeat myself 5.5 and 5.4 is better than claude-sonnet5) but for the way it speaks and talk, 5.6 is on another level
using words like "faitful" "reconstruction"

im not a big fan of
• Ran python -m pip download --no-deps --dest C:\tmp mc10-tools==0.9

but its not like it used a command that wouldnt work

actually , Ive put put 5.6 to the test, and it actually used a library and tool --- while 5.5 didnt used a tool and it didnt tried to reinvent the wheel and analyse the a literal .c10 binary

so very good!

I honestly was so disapointed at 5.2 and 5.3 (but I like the way it tries to talk with you regarding task), 5.4 was truely something else , but I understand why, because the older models didnt really compete with the other models despite that I saw the pontential because for using windows powershell at the time only openai done it correctly

so yeah very good update, considering things are now faster ---- and I think it choses things that are good and things that the users would be happy (you dont need to write a C++ compiler for everything!)

[eah3699]

yes i found the instruction following and using tools more efficiently and actually reading and understanding intent and meaning is much better. It finally runs like GPT but with codex. Huge improvement. In the past 5 hours i made more progress than i have in a week because it also handles sub-agents very well too.

[Mahkhmood9]

Honestly I think the new naming changes is great, I think the users should refain from use the bigger models all the time , when they can be just be happy with the smaller models.
granted I think 5.5 was trained with poor quiality of data (not really, but compared to now) (which really shows how far things can still can get even from the point of today) and to think a lot of the data used is actually synthetic and hypothesized, I also think its a great thing that the models released now , also preform bad on some tasks and evals. becauses errors to some degree mean creativety.

and to be honest, I am sure in a lab testing openai can achieve even 80% where Sol 5.6 eached on evals 60%... (not-or 0-shot, or very expensive even in lab setting)

and also I guess again , I hope people try to expirement with the smaller models first and try to actually adjust the smaller models like tera or luna, and not go full power (because unlike other companies, if you choose the bigger models, they do not degrade your service to the same degree more than 30%-10% , which is not 65%)

I personally wasnt a fan of Claude Code, because it took 5 seconds to understand at december 2025 that their documentation is cheotic (more than my writing style yes) and so much stuff in their documentaiton gets aboanded in like 3 days
(and the same I cant say on Apple or Microsoft btw, but they are commiting this kind of thigns too) I personally think Microsoft and Nvidia and Intel are the kings of long term support (and JEDEC RAM , other open standards, and pci-express - you see the pattern where things attempt to be more minimalist with hardware)

as per Klaus Claude, recently that launch a feature called ULTRATHINK, and it would literally will launch like 105subagents or whatever, and it would burn into your 5-houry credit in 3 prompts on a small [emtypical-vibe-coded small code] , how dare they release this to users? also burning all of this compute when they clearly cant serve as things going now

as per subscribtion side , I also think if I paid in 2024 for few months in a faw for a claude subscribtion , I do hope to get basic prompt ablity ,, because once I am not a paying customer and I became a free users--- I I should become like free_paid_in_the_past and get a very basic preview, but instead I used other third party services - which lets be honest poision their data -- and anthropic doesnt enjoy the same kind of luxury of most people using MicrosoftCoppilot or Codex, and a lot of the things regarding how messages are handled happen quickly openai even changed the their standards regarding text/agent-chatting -- anthropic has a lot of data from a lot of third party apps --- which I assume they can literally use llm based apporch to even clean the dataset (find start, and end goal --- or if the user achieved what he wanted) as per raw chat, I actually thing it can get creative, but I think it get posion with a massive context window you already load before a you start a new chat (system card? is that that?) I actually think that adding a lot of guildlines and safety or whatever are good for the models, because it adds a lot of complexity on how the models themselves are going to desgined and I think its part of the reason it got driven so fast in the last two years -- because it does degrade preformance and it forces team to expirement more and more (but thats just a theory of mine because I believe you should break your head with a lot of limitation) lets real there is more than just matrix multipications , every itiration, would provide much more rather than labeling it as a data problem and pass it to the cpu,
despite gpu are faster than gpu 5000 times but so is camera sensors which can get to 5000 fps and a lot more, because we are limited by electronic gates right, and not expousurer ---- but on the other hand you can find great artistic value if you used on film camera -- as things start now the age of the internet is about 30 years -- as things stands about harddrives/ssd you dont trust them to last more than 5-10 years despite being one of the most resiliant things you can buy hardware wise --- I dont think I see any ssd failure on any iphone or android device so we can consider it 101% perfect as the 2% are some weird physical thing that happen to it,, but think how we hobbits folk look at code we written very simple code , and we keep getting c compiler errors - when it wasnt happening when you were at the vibe when you were making the software - also forggeting what you really wanted to do , in that case you get a lot less distruction, as precious and gracious doing things with your hands is, and having other people do craft with their hands,, and not be a mad hatter(as per the the trade of hatmaking) because of heavy metals you inhaled due to the craft

[Mahkhmood9]

you are having issues, can you please share the prompts and things you are trying to achieve --- and I will respond and try to grill you (in a friendly way --- to see how things should be broken down? and that you get mroe efficent) (yes my few years I was living in the netherlands start showing here haha )

[Mahkhmood9]

I tell you what you wont be disappointed with the smaller models - I dont think you will tell the difference, despite what you choose is what you get --- (that cant be said on other companies)

[eah3699]

I do a lot of long horizon heavy context work. I noticed a difference once I did Light and Medium. High performed well but wasn't as efficient as xHigh surprisingly. Took longer but actually used less tokens because it was more efficient. But definitely not to use if doing quick lookups online or something. It burned through 41% context windows once when i wasn't paying attention and I asked it a simple question haha

[Mahkhmood9]

I do a lot of long horizon heavy context work. I noticed a difference once I did Light and Medium. High performed well but wasn't as efficient as xHigh surprisingly. Took longer but actually used less tokens because it was more efficient. But definitely not to use if doing quick lookups online or something. It burned through 41% context windows once when i wasn't paying attention and I asked it a simple question haha

@eah3699
Look, I personally not a big fun of ai becoming make a 4000 words prompt , I think there is some level of techinicalies you hsould also know and understand yourself.. this is not complex chemical synthesis or a complex way to manufactor car engines and reduce cost (also see how they are costly errors and hard lessons everywhere)
I personally think, developing is the art of craft --- im not sure thats the way you should be using coding agents --- and I think there is a beauty in small consive prompts...

at the rate and jump between 5.6 from 5.5 . I honestly recomend you to avoid "Sol" and use terra --- I recomend you force yourself developing the same masscle memory as with getting the same achievements.

I get it does depend in which tech indurstry you are and what are you task.... I think a lot of the issue people having are because - a lot of people hold a job title and they dont know what are they doing... and look I am not against it, people have morgadges or whatever - and economically speaking having people unemployed is unacceptable I just hope people would been more graceful for what they got.