alt Hacker NewsGPT-5.3-Codex
Comments
I think Anthropic rushed out the release before 10am this morning to avoid having to put in comparisons to GPT-5.3-codex!
The new Opus 4.6 scores 65.4 on Terminal-Bench 2.0, up from 64.7 from GPT-5.2-codex.
GPT-5.3-codex scores 77.3.
,,GPT‑5.3-Codex is the first model we classify as High capability for cybersecurity-related tasks under our Preparedness Framework , and the first we’ve directly trained to identify software vulnerabilities. While we don’t have definitive evidence it can automate cyber attacks end-to-end, we’re taking a precautionary approach and deploying our most comprehensive cybersecurity safety stack to date. Our mitigations include safety training, automated monitoring, trusted access for advanced capabilities, and enforcement pipelines including threat intelligence.''
While I love Codex and believe it's amazing tool, I believe their preparedness framework is out of date. As it is more and more capable of vibe coding complex apps, it's getting clear that the main security issues will come up by having more and more security critical software vibe coded.
It's great to look at systems written by humans and how well Codex can be used against software written by humans, but it's getting more important to measure the opposite: how well humans (or their own software) are able to infiltrate complex systems written mostly by Codex, and get better on that scale.
In simpler terms: Codex should write secure software by default.
Something that caught my eye from the announcement:
> 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
I'm happy to see the Codex team moving to this kind of dogfooding. I think this was critical for Claude Code to achieve its momentum.
I remember when AI labs coordinated so they didn't push major announcements on the same day to avoid cannibalizing each other. Now we have AI labs pushing major announcements within 30 minutes.
I've always been fascinated to see significantly more people talking about using Claude than I see people talking about Codex.
I know that's anecdotal, but it just seems Claude is often the default.
I'm sure there are key differences in how they handle coding tasks and maybe Claude is even a little better in some areas.
However, the note I see the most from Claude users is running out of usage.
Coding differences aside, this would be the biggest factor for me using one over the other. After several months on Codex's $20/mo. plan (and some pretty significant usage days), I have only come close to my usage limit once (never fully exceeded it).
That (at least to me) seems to be a much bigger deal than coding nuances.
Actually kind of excited for this. I've been using 5.2 for awhile now, and it's already pretty impressive if you set the context window to "high".
Something I have been experimenting with is AI-assisted proofs. Right now I've been playing with TLAPS to help write some more comprehensive correctness proofs for a thing I've been building, and 5.2 didn't seem quite up to it; I was able to figure out proofs on my own a bit better than it was, even when I would tell it to keep trying until it got it right.
I'm excited to see if 5.3 fairs a bit better; if I can get mechanized proofs working, then Fields Medal here I come!
Many are saying codex is more interactive but ironically I think that very interactivity/determinism works best when using codex remotely as a cloud agent and in highly async cases. Conversely I find opus great locally, where I can ram messages into it to try to lever its autonomy best (and interrupt/clean up)
I've been listening to the insane 100x productivity gains you all are getting with AI and "this new crazy model is a real game changer" for a few years now, I think it's about time I asked:
Can you guys point me ton a single useful, majority LLM-written, preferably reliable, program that solves a non-trivial problem that hasn't been solved before a bunch of times in publicly available code?
Terminal Bench 2.0
| Name | Score |
|---------------------|-------|
| OpenAI Codex 5.3 | 77.3 |
| Anthropic Opus 4.6 | 65.4 |> Using the develop web game skill and preselected, generic follow-up prompts like "fix the bug" or "improve the game", GPT‑5.3-Codex iterated on the games autonomously over millions of tokens.
I wish they would share the full conversation, token counts and more. I'd like to have a better sense of how they normalize these comparisons across version. Is this a 3-prompt 10m token game? a 30-prompt 100m token game? Are both models using similar prompts/token counts?
I vibe coded a small factorio web clone [1] that got pretty far using the models from last summer. I'd love to compare against this.
When 2 multi billion giants advertise same day, it is not competition but rather a sign of struggle and survival. With all the power of the "best artificial intelligence" at your disposition, and a lot of capital also all the brilliant minds, THIS IS WHAT YOU COULD COME UP WITH?
Interesting
Do software engineers here feel threatened by this? I certainly am. I'm surprised that this topic is almost entirely missing in these threads.
I would love to see a nutritional facts label on how many prompts / % of code / ratio of human involvement needed to use the models to develop their latest models for the various parts of their systems.
The behind the scenes on deciding when to release these models has got to be pretty insanely stressful if they're coming out within 30 minutes-ish of each other.
Using opus 4.6 in claude code right now. It's taking about 5x longer to think things through, if not more.
I want to recompile a Rust project to be f32 instead of f64.
Am I better off buying 1 month of Codex, Claude, or Antigravity?
I want to have the agent continuesly recompile and fix compile errors on loop until all the bugs from switching to f32 are gone.
Both Opus 4.6 and GPT-5.3 one shot a Gameboy emulator for me. Guess I need a better benchmark.
> our team was blown away > by how much Codex was able > to accelerate its own development
they forgot to add “Can’t wait to see what you do with it”
> GPT‑5.3-Codex was co-designed for, trained with, and served on NVIDIA GB200 NVL72 systems. We are grateful to NVIDIA for their partnership.
This is hilarious lol
For those who cared:
GPT-5.3-Codex dominates terminal coding with a roughly 12% lead (Terminal-Bench 2.0), while Opus 4.6 retains the edge in general computer use by 8% (OSWorld).
Anyone knows the difference between OSWorld vs OSWorld Verified?
Did they post the knowledge cutoff date somewhere
AI designed websites are so easy to spot that I need to actively design my UI so that it doesn't look AI
I think models are smart enough for most of the stuff, these little incremental changes barely matter now. What I want is the model that is fast.
I'm having a hard time parsing the openai website.
Anyone know if it is possible to use this model with opencode with the plus subscription?
Gotta love how the game demo's page title is "threejs" – I guess the point was to demo its vibe-coding abilities anyway, but yea..
Anthropic mostly had an advantage in speed. It feels like with a 25% increase in speed with Codex 5.3, they are now losing that advantage as well.
I never really used Codex (found it to slow) just 5.2, which I going to be an excellent model for my work. This looks like another step up.
This week, I'm all local though, playing with opencode and running qwen3 coder next on my little spark machine. With the way these local models are progressing, I might move all my llm work locally.
It's so difficult to compare these models because they're not running the same set of evals. I think literally the only eval variant that was reported for both Opus 4.6 and GPT-5.3-Codex is Terminal-Bench 2.0, with Opus 4.6 at 65.4% and GPT-5.3-Codex at 77.3%. None of the other evals were identical, so the numbers for them are not comparable.
I find it very, very interesting how they demoed visuals in the form of the “soft SaaS” website and mentioned how it can do user research. Codex has usually lagged behind Claude and Gemini when it comes to UX, so I’m curious to see if 5.3 will take the lead in real world use. Perhaps it’ll be available in Figma Make now?
I think this announcement says a lot about OpenAI and their relationship to partners like Microsoft and NVIDIA, not to mention the attitude of their leadership team.
On Microsoft Foundry I can see the new Codex 4.6 model right now, but GPT-5.3 is nowhere to be seen.
I have a pre-paid account directly with OpenAI that has credits, but if I use that key with the Codex CLI, it can't access 5.3 either.
The press release very prominently includes this quote: "GPT‑5.3-Codex was co-designed for, trained with, and served on NVIDIA GB200 NVL72 systems. We are grateful to NVIDIA for their partnership."
Sounds like OpenAI's ties with their vendors are fraying while at the same time they're struggling to execute on the basics like "make our own models available to our own coding agents", let alone via third-party portals like Microsoft Foundry.
Interesting that this was released without a prior GPT-5.3 release. I wonder if that means we won't see a GPT-5.3?
gpt-5.3-codex isn't available on the API yet. From TFA:
> We are working to safely enable API access soon.
GPT-5.2-Codex was so cool at price/value rate, hope 5.3 will not ruin the race with claude
May AI not write the code for me.
May I at least understand what it has "written". AI help is good but don't replace real programmers completely. I'm enough copy pasting code i don't understand. What if one day AI will fall down and there will be no real programmers to write the software. AI for help is good but I don't want AI to write whole files into my project. Then something may broke and I won't know what's broken. I've experienced it many times already. Told the AI to write something for me. The code was not working at all. It was compiling normally but the program was bugged. Or when I was making some bigger project with ChatGPT only, it was mostly working but after a longer time when I was promting more and more things, everything got broken.
Having used codex a fair bit I find it really struggles with … almost anything. However using the equivalent chat gpt model is fantastic. I guess it’s a matter of focus and being provided with a smaller set of code to tackle.
That was fast!
I really do wonder whats the chain here. Did Sam see the Opus announcement and DM someone a minute later?
Anyone remember the dot-com era when you would see one provider claim the most miles of fibre and then later that week another would have the title?
Funny that this and Opus 4.6 released within minutes of each other. Each showing similar score improvements. Each claiming to be revolutionary.
I've been using 5.2 the way they're describing the new use case for 5.3 this whole time.
is so fun that the two releases used almost completely non-overlapping benchmarks!
How'd they both release at the same time? Insiders?
At first try it solved a problem that 5.2 couldn't previously.
Seems to be slower/thinks longer.
So can I use this from Opencode? Because Anthropic started to enforce their TOS to kill the Opencode integration
Whats interesting to me is that these gpt-5.3 and opus-4.6 are diverging philosophically and really in the same way that actual engineers and orgs have diverged philosophically
With Codex (5.3), the framing is an interactive collaborator: you steer it mid-execution, stay in the loop, course-correct as it works.
With Opus 4.6, the emphasis is the opposite: a more autonomous, agentic, thoughtful system that plans deeply, runs longer, and asks less of the human.
that feels like a reflection of a real split in how people think llm-based coding should work...
some want tight human-in-the-loop control and others want to delegate whole chunks of work and review the result
Interested to see if we eventually see models optimize for those two philosophies and 3rd, 4th, 5th philosophies that will emerge in the coming years.
Maybe it will be less about benchmarks and more about different ideas of what working-with-ai means