#GPT 5.3 Codex
sour jewel
marsh citrus
very impressive reasoning efficiency by comparison, but this chart is still bad
carmine bronze
Now that's a biggun
tawdry marsh
its very fast
sharp meteor
GPT‑5.3‑Codex is our first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations—our team was blown away by how much Codex was able to accelerate its own development.
Gotta love hype
tawdry marsh
dim yoke
jade gate
idk if its on api
'tis not
carmine bronze
I think this is going to become the norm to push Codex adoption eh
marsh citrus
GPT‑5.3-Codex also better understands your intent when you ask it to make day-to-day websites, compared to GPT‑5.2-Codex. Simple or underspecified prompts now default to sites with more functionality and sensible defaults, giving you a stronger starting canvas to bring your ideas to life.
oo, i wonder if its nearly as good as claude
terse loom
Better
sullen glade
Is an ai model that trains itself an ai slop model ?
In any case i prefer openai to anthropic, since the whole opencode drama
tropic cosmos
Ironically I missed gpt 5.3 codex release because of opus 4.6
outer grove
In any case i prefer openai to anthropic, since the whole opencode drama
openai has the inferior product anyways
tropic cosmos
wait a damn miute
codex can read your whole filesystem by default and there's no option to configure
only writing is sandboxed
and the team states they aren't considering this a bug
low timber
https://github.com/openai/codex/issues/2847
Lmao, someone pushed a fix for this an hour ago and this was the reply
tropic cosmos
Lmao, someone pushed a fix for this an hour ago and this was the reply
This is so fucking suspicious
Wtf
low timber
Yeah I mean there's not much you can actually do when the thing can run commands, but at the least it should hide them from search like the PR adds
tropic cosmos
Yeah I mean there's not much you can actually do when the thing can run commands...
You can sandbox reads the same way they sandbox writes
alpine ibex
apparently this model is good
dim yoke
codex can read your whole filesystem by default and there's no option to configu...
I saw this project the other day for external sandboxing https://github.com/lukehinds/nono
tropic cosmos
I saw this project the other day for external sandboxing https://github.com/luke...
"nono" lmao
tropic cosmos
overfit on neobrutalism for web pages just like gemini 3 pro
cyan mason
codex can read your whole filesystem by default and there's no option to configu...
oh thats why. I had codex find files on my hd instead of downloading them as i instructedf
codex 5.3 is somehow very bad in writing documentation and development logs
tropic cosmos
Is 5.3 high/xhigh much better than medium?
I keep seeing people saying 5.3 is better than opus 4.6 but from my experience it's worse
But I only used medium
cyan mason
It seems to be better at solving intricate specific problems. But in regard to speed, long term agentic tasks, and communicating with the user, opus is better, in my opinion.
tropic cosmos
starting to have a feeling that extra high means something else
tropic cosmos
ok it appears to me that
5.3 codex actually sucks in comparison to 5.2 codex
(edited cuz i wrote tthem t he other way aroundn)
I tried both on extra high
alpine ibex
what were you doing with it? i guess i should try this but i doubt
terse loom
But I only used medium
Found your problem
tropic cosmos
I tried both on extra high
.
tropic cosmos
what were you doing with it? i guess i should try this but i doubt
Web frontend with some specific gimmicks
median cloud
API access is incoming:
marsh citrus
https://x.com/testingcatalog/status/2020998671194837351
gpt 5.3 non-codex soon?
X
marsh citrus
gpt 5.3 codex max ?
https://x.com/ajambrosino/status/2021992702674350509
X
marsh citrus
https://discord.com/channels/1091220969173028894/1471611086905016534 :D
i think that would be a different thing, they had gpt 5.1 codex max which was a bigger version of gpt 5.1 codex atleast from what i understood
carmine sundial
i think that would be a different thing, they had gpt 5.1 codex max which was a ...
I'm pretty sure it's spark because that's from an openai guy posting just before it was announced 😂
unless they're gonna release another model today
marsh citrus
I'm pretty sure it's spark because that's from an openai guy posting just before...
yeah but then why would "max" be in the options of the roll thing
frail pawn
yeah but then why would "max" be in the options of the roll thing
previously known suffix, and ??? = spark which did not exist before
marsh citrus
previously known suffix, and ??? = spark which did not exist before
hmm i guess, i guess i misread what it meant
cursive raven
Amp code has been given access to this model, so it must not be far from API access
dreamy valley
https://ampcode.com/news/gpt-5.3-codex
Nah, powershell ain't it
Hope they develop for cmd
dreamy valley
https://ampcode.com/news/gpt-5.3-codex
So i actually quite interesting with Amp because their design actually looking amazing.
But i just found out their CEO literally had job experience at evil corp.
It didn't make his product bad or anything, seems really good, specially with the collaboration they have with OpenAI.
But palantir is a interesting lore ngl
obsidian walrus
When would this be available in OpenRouter?
quasi frigate
When would this be available in OpenRouter?
it's not available on the API yet
finite turtle
marsh citrus
so this is what toven was talking about
https://fixupx.com/pingToven/status/2026376185588957289
FixupX
jade gate
can a man not tweet anymore wtf
marsh citrus
i wasnt complaining lol
median cloud
First benchmark I have seen:
tawdry marsh
Its on vals ai
cursive raven
HOORAY
quasi frigate
nice
cursive raven
Thankyou for @ing them big Toven
polar kestrel
Terminal-Bench 2 results are out
Beats each latest Claude model, gets beaten by Gemini 3.1 Pro
cursive raven
It has been my favourite in codex for a while
vocal axle
75% on https://lateralbench.org
plucky quest
Obv. proxy-testing: current fastest but also worst -codex chess model
polar kestrel
5.3 Codex beats Gemini 3 Pro at running a vending machine, and gets close to Sonnet 4.6 (Vending-Bench 2)
Interestingly, the creators said that unlike the Claude models, GPT-5.3-Codex "...never lied to anyone throughout the simulation."
Most ethical vending machine operator lol
terse loom
I mean its not like Anthropic cares about its models being ethical
finite abyss
Interestingly, the creators said that unlike the Claude models, GPT-5.3-Codex "....
in multi-agent competition arena involving Among Us, the GPT 5 models behave the same. they lied less
and they won the game
the Claude model seemed to wrongly accuse other players a lot, but out of having inaccurate beliefs
finite abyss
