GPT-5.3-Codex - Expanding Codex to the full spectrum of computer work
by•
Advances the frontier of coding and computer work. SOTA on SWE-Bench Pro (57%) and OSWorld (64%). Features mid-task steerability (interact while it works), 25% faster speeds, and "High" capabilities in cybersecurity.
Replies
Flowtica Scribe
Hi everyone!
GPT-5.3-Codex is here.
The benchmark jumps are impressive (especially OSWorld going from ~38% to 64%), but I found this specific detail in the announcement most interesting:
The team used early versions of the model to debug the training run, manage deployment, and diagnose test results. It basically accelerated its own development.
Codex is becoming a broader productivity agent that can handle complex workflows end-to-end.
It is available now for paid ChatGPT plans, everywhere you can use Codex: the app, CLI, IDE extension and web. API on the way.
Migma AI
Cool addition!
The practical difference here is execution + iteration: it can take a task, make changes, run/validate, and refine without needing a new prompt for every bump. The frequent status updates and mid-course steering are what made it useful for real repo work (refactors, failing tests, debugging). I still review diffs carefully—especially anything touching auth/security—but it’s a legitimate productivity boost compared to earlier Codex versions.
Kilo Code
@OpenAI announced the model and it's live on @Kilo Code within minutes. Compared to prior Codex models, it's more token-efficient and approximately 25% faster.
It feels like they unified two of its most powerful research tracks: the frontier code generation of the Codex line and the elite professional reasoning of the GPT-5.2 "Thinking" series.
LFG!