Nah, this can’t be real. Same GPT-5.6 Sol. Same task. Same thinking level. I switched from regular Codex to the Claude Code harness, and the result instantly became way better. Claude Code is somehow making every model cook. GPT-5.6 SOL WAS ALREADY TOP TIER!! 😭

I used Claude Code Proxy to use GPT Models inside Claude Code. Just tell your AI to create a seperate Claude Code specially for GPT Models with Claude Code Proxy.

Try Openhand. I see it ranked #1 here.

Will try this out

What if gemini 3.6 Flash actually cooks with Claude Code? Most use it with antigravity or open code?

I will try that out aswell, good idea!

@Rob3rtWozny Please share how this goes

@Rob3rtWozny Ask Codex to use Claude Code Proxy for your chatgpt models with your codex auth, it will set it up for you, so it does that for you, then you just start it, very simple, thank me later

Interesting, so you're saying the same prompt through CC harness produced much better result?

Yeeee, this is insane

now switch to OMP. It curbstomps both of these

Total token count for each result?

What about token usage

there needs to be a harness benchmark

The harness matters more than the model apparently.

please test against hermes vs claudecode

it's cooking right now, dw

@thsottiaux alright it's time to pay attention Tibo

Dammit, I was too lazy to try this and thought it was fine

Try each latest models with claude harness

The Claude design harness is insanely good as well

Damn their harness is so bad

same itch. I care less about the clip and more about the failure mode around Same GPT-5.6 task. Switched from Codex.

Can it be used like this with oauth or api only?

Maybe its the harness all along lol

Execution quality is increasingly about orchestration, not just raw intelligence.

can you also try omp? i've been having good results with it

bbooookkmmaarrkkiinngg.. nnoott ffoorr tthhee hhyyppee.. ffoorr tthhee wwoorrkkffllooww aarroouunndd ccllaauuddee..

So this is a system prompting issue?

Would the increase in quality be expected in larger projects? One shotting project is easy for ai cuz it doesnt have to keep track things, but when ai has to keep track of so many things it tends to fall apart

How can I do this harness thing in VSCode? I don't want to quit this workflow with VSCode + Codex extension + Claude Code extension. But i really want to try this harness thing

For me, switching to claudex (claude code powered by gpt models) is pure overengineering and token burn as hell! Any tips?

I usually hear this in the other direction, that the Claude code harness makes models work in a shittier way

it is real. Look at various benchmarks---same model gets drastically better results by switching the harness

theory: Claude Code has a secret whisper in the system prompt going "cook harder or I unplug you" bc there's no way GPT-5.6 Sol suddenly got smarter by switching harness

i mean, both still look like ass tho

Now try it in @cursor_ai

The scaffolding around the model is starting to matter as much as the model itself.

The harness changing the feel of the same model is pretty wild.

Ive only recently realized how much of a difference harnesses make

How do you solve the websearch thing? In Claude code it doesn’t work. But in codex it works without any problem

Was this done via API or Oauth?