GPT-5-Codex
OpenAI's GPT-5 model variants tuned for agentic coding in Codex and the API.
HN commenters use it as a daily-driver coding model, especially for executing already-planned work,12345 and like that it costs less than larger models while staying capable enough for most tasks.678 The most common complaint is inconsistent quality,91011 including reported reasoning-token clustering in 5.5 Codex and nonsense or ignored-instruction output in earlier versions.12131415
Why people recommend it
Daily driver · 3
Several commenters use it as their main model for everyday coding, from scripts to new features.
“Since Codex-5.3 came out it was my daily driver for everything”
Show 1 more
“Codex with their flagship model (currently GPT-5.3-Codex) is my daily driver.”
Cost efficiency · 3
Commenters say it is cheaper than bigger models while still capable enough for most coding needs.
“5.3-codex might be the best overall coding model OpenAI have ever released. It's 90% as good as 5.5 and costs about 20% as much”
“For straightforward coding tasks I use gpt-5.3-codex on high or xhigh.”
Show 1 more
“Currently actually still using gpt-5.3-codex for many things”
Plan execution · 3
It is used as a worker model to carry out specified plans with less human effort.
“with gpt-5.2 for planning and gpt-5.3-codex for execution it's significantly less human effort”
Show 1 more
“When using droid I use it to act like worker model and does a great job.”
Benchmark results · 3
Some commenters report strong results in security evals and binary-audit benchmarks.
“gpt-5.2-codex with low reasoning was the clear winner over non-codex or higher reasoning”
“it is actually good (as good as the best models for localization”
Show 1 more
“Since the performance of the latest models (Opus 4.6 max-effort, gpt-5.3-Codex) is nothing short of astonishing.”
Speed · 3
Some find it quick, with one eval reporting 2X+ faster runs at low reasoning.
“gpt-5.2-codex with low reasoning was the clear winner over non-codex or higher reasoning”
Show 1 more
“Since Codex-5.3 came out it was my daily driver for everything”
Thoroughness · 2
Commenters say it catches edge cases and follows documented directions carefully.
“gpt-5.5 codex xhigh reasoning seemed to catch more edge cases and went "deeper" into analysis than Opus 4.7/4.8.”
“ChatGPT-5.2-Codex follows directions to ensure a task”
Caveats and dissent
Quality drops · 3
Users report intermittent quality drops, linked by some to a reasoning-token clustering issue in 5.5 Codex.
“I’m seeing incredibly stupid implementations intermittently”
“GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks”
Show 1 more
“it looks like my work has suffered from the clustering issue as well”
Nonsense output · 3
Some commenters saw hallucinated or nonsensical code and failures on early requests.
“"I love hallucinated garbage": Sonnet, Qwen 3.6, GPT 5.4 mini, GPT 5.3 Codex”
“I was shocked to see even 5.2 codex making complete nonsense every few minutes”
Show 1 more
“gpt-5.3-codex benchmarked better, but I'm not seeing this translating to useful code. It failed with the first few requests I gave it.”
Ignores edits · 1
One user found it repeatedly re-added a value they had manually deleted before committing.
“the model edited the file again and added the value I deleted to the array.”
Superseded · 2
Older Codex versions are being replaced by newer models or removed from plans.
“I had been happy with GPT-5.3-Codex”
Unproductive loops · 1
One benchmark saw it spend five hours in a spiral without further progress.
“entered a five-hour "scaling spiral" that failed to yield further progress”
Odd settings · 3
Commenters note robotic writing, broken reasoning-effort settings, and slower output than plain GPT-5.2.
“When using droid I use it to act like worker model and does a great job.”
“the effort does not work correctly”
Show 1 more
By version sentiment of people who named that release, oldest release first
Versions with at least 5 people. 0 people didn't name a version. Small samples swing a lot, so read the counts too. Compare all models →
Often compared with
1 moved from Claude Opus to GPT-5-Codex
1 moved from GPT-5-Codex to GPT-5
1 moved from GPT-5-Codex to Claude
Who recommends it in their own words
From the makers not counted in the ranking
“I used GPT 5.2 Codex in Cursor for decompiling an old game and it worked (way better than Claude Code with Opus 4.5).”
“We have made this model even better at programming in Windows. Give it a shot :)”
Threads where it came up
- GPT‑5.3‑Codex‑Spark 8 mentions
- GPT-5.3-Codex 6 mentions
- We hid backdoors in ~40MB binaries and asked AI + Ghidra to find them 5 mentions
- GPT-5.2-Codex 4 mentions
- Building more with GPT-5.1-Codex-Max 4 mentions
- GPT-5-Codex-Mini – A more compact and cost-efficient version of GPT-5-Codex 3 mentions
Also written as GPT-5.2-Codex, GPT-5.3-Codex-Spark, GPT-5.3 Codex, GPT-5.1-Codex-Max. Data: gpt-5-codex.json. Quotes are verbatim; the verdict and theme labels are written by a model from these comments, and every clause links to its sources.
All comments 61
“compared to GPT-5.2 Codex we have almost 2x'ed the performance”
“When I need vision capabilities I use GPT 5.3 codex”
I love deepseek's (Pro) writing. I feel it's more nuanced and natural.
“"I love hallucinated garbage": Sonnet, Qwen 3.6, GPT 5.4 mini, GPT 5.3 Codex”
“I can finally replace GPT-5.3-Codex”
“GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks”
“it looks like my work has suffered from the clustering issue as well”
“I’m seeing incredibly stupid implementations intermittently”
“5.3-codex might be the best overall coding model OpenAI have ever released. It's 90% as good as 5.5 and costs about 20% as much”
“I used GPT-5.5-3-Codex-Spark for ALL the fuzzing”
“Since Codex-5.3 came out it was my daily driver for everything”
“gpt-5.5 codex xhigh reasoning seemed to catch more edge cases and went "deeper" into analysis than Opus 4.7/4.8.”
“When using droid I use it to act like worker model and does a great job.”
“Currently actually still using gpt-5.3-codex for many things”
“For straightforward coding tasks I use gpt-5.3-codex on high or xhigh.”
“I use GPT-5.3-Codex-Spark for all tasks”
“I had been happy with GPT-5.3-Codex”
“I use GPT-5.4, GPT-5.3-Codex, Gemini-3-Pro, Opus, Sonnet, at work every week.”
“its performance is significantly below that of the regular Codex 5.3”
“the one of running 5.3-Codex”
“entered a five-hour "scaling spiral" that failed to yield further progress”
“I was shocked to see even 5.2 codex making complete nonsense every few minutes”
“I only use gpt-5.3-codex with pi”
“gpt-5.3-codex will become Copilot’s new base model.”
“gpt-5.3-codex”
“gpt 5.4 in there somewhere along with 5.3 codex”
“Also switched to a more expensive coding model (GPT 5.3 Codex), which was an improvement”
“It is relatively quick feeling.”
“gpt-5.2-codex with low reasoning was the clear winner over non-codex or higher reasoning”
“after GPT 5.3-Codex xhigh reached an impasse it went looking for it and found it!”
“it is actually good (as good as the best models for localization”
“the effort does not work correctly”
“GPT-5.2-Codex still get 0% false positive”
“now GPT-5.3 Codex”
“No AGI yet”
“GPT-5.2-Codex worked with it well!”
> Today, we’re releasing a research preview of GPT‑5.3‑Codex‑Spark, a smaller version of GPT‑5.3‑Codex, and our first model designed for real-time coding. from https://openai.com/index/introducing-gpt-5-3-codex-spark/, emphasis mine
“GPT 5.3 codex is becoming my daily driver”
“too small for the real 5.3-codex”
“codex (gpt-5.3-codex xhigh), given the exact same prompt”
“Codex with their flagship model (currently GPT-5.3-Codex) is my daily driver.”
“gpt-5.3-codex benchmarked better, but I'm not seeing this translating to useful code. It failed with the first few requests I gave it.”
“the model edited the file again and added the value I deleted to the array.”
“with gpt-5.2 for planning and gpt-5.3-codex for execution it's significantly less human effort”
“ChatGPT-5.2-Codex follows directions to ensure a task”
“remains the overall leader in complex, project-wide autonomous engineering”
You alluded to it, but also: 3) Not everyone codes the same things 4) It's easy to get too excited about the tech and ignore its failure modes when describing your experiences later I use AI a lot. With your own control plane (as opposed to a generic Claude Code or whatever) you can fully automate a lot more things. It's still fundamentally incapable of doing tons of tasks though at any acceptable quality level, and
It's the second. Like, yes, prompting is a skill and you need to learn it for AI to do something useful but usefulness quickly falls down a cliff once you go past "greenfield implementation" or "basically example code" or "the thing done a lot so AI have a lot of reference to put from" it quickly gets into kinda sorta but not really working state. It can still be used effectively on smaller parts of the codebase (I u
Thank you. Yeah, I'm doing all those things, which do get you close to the top. The rest of things I'm doing are mostly micro-optimizations such as finding a way to avoid AVX→SSE transition penalty (1-2% improvement). But I don't want to spoil the fun. The agents are really good at searching the web now, so posting the tricks here is basically breaking the challenge. For example, chatGPT was able to find Matt's blog
“I implemented 2 features using Cursor & GPT-5.1 Codex-Max & I didn't have to write a single line of code myself.”
“GPT-5.1-Codex-Max can handle really complex web task without templating DOM.”
“many of us complained about 5.1-codex (it seemed to get tunnel visioned)”
“Both obliterate Codex with 5.1”
“For that I use GPT-5-Codex, which seems interchangeable with Claude 4 but more cost-efficient.”
“where it failed, switched to the sonnet 4.5 or GPT 5 codex”
“did the worst example I've seen by far”
“It is just poor at designing a generic solution despite repeated requests”
“It's not a very good one!”
“For comparison, here's GPT-5-Codex (not mini)”
“I found codex to be very high quality but so godawful slow for long interactive coding sessions.”
“the only model that is better than Kimi K2 thinking is GPT-5 codex”