GPT-5.3-Codex — The Self-Improving Agentic Coding Model
New industry highs on SWE-Bench Pro / Terminal-Bench / OSWorld / GDPval. Real-time steering during execution. The first Codex model used to train itself.
Timeline & Positioning
OpenAI released GPT-5.3-Codex on 2026-02-05 as the successor to 5.2-Codex. It advances both the frontier coding performance of 5.2-Codex and the reasoning / professional knowledge of GPT-5.2 in one model, while being 25% faster overall. Available across all Codex surfaces: Codex macOS app, CLI, IDE extension, and Codex on the web. Available to paid ChatGPT plan users; API access rolls out as OpenAI completes safety reviews.
Four SOTA Benchmarks
SWE-Bench Pro 56.8% (industry-leading), Terminal-Bench 2.0 77.3%, OSWorld-Verified 64.7%, and a new high on GDPval. SWE-Bench Pro is a tougher repo-rewrite benchmark than the more familiar SWE-Bench Verified (where 5.2-Codex already hit 80%). Terminal-Bench measures end-to-end terminal agent tasks; OSWorld measures OS-level GUI tasks.
Real-Time Steering and Self-Improvement
Unlike prior models you can steer and interact with GPT-5.3-Codex while it's working — watch progress in the Codex macOS app, ask questions, provide feedback, or correct course mid-task without breaking the workflow. OpenAI also disclosed that the Codex team used early 5.3-Codex versions to debug its own training, manage deployment, and diagnose evaluations — making it the first Codex model that was instrumental in creating itself.
Choosing Between GPT-5.5 / Opus 4.7 / GPT-5.3-Codex
GPT-5.5 is OpenAI's general-purpose flagship — strong across coding, research, writing, and tool use. GPT-5.3-Codex is the coding-optimized branch of that family, SOTA on long-horizon coding tasks. Claude Opus 4.7 leads on Python/Django refactor depth and long-context RAG. Short version: pick 5.3-Codex for pure long-horizon coding, 5.5 for mixed code+research, Opus 4.7 for deep Python refactors.
FAQ
Can I access GPT-5.3-Codex from China directly?
OpenAI's official channels are restricted in mainland China. QCode.cc provides transparent unified API access to GPT-5.5 and the Codex family from inside China with a single subscription — no self-hosted proxy required.
Should I upgrade from 5.2-Codex immediately?
If your workload is long-horizon multi-step agents (end-to-end PRs, cross-file refactors, terminal tasks), yes — benchmarks are clearly stronger and it's 25% faster. For short completions the gap is smaller; you can wait.
How does real-time steering work in practice?
While the model is executing a long task in the Codex macOS app or CLI, you can type new instructions or context into the same session. The model incorporates the new signal without invalidating its current reasoning tree — the key interaction upgrade over 5.2-Codex.
Access GPT-5.3-Codex / GPT-5.5 / Opus 4.7 through QCode
Transparent access to all major coding models from inside China with unified billing — no need to buy plans from each vendor separately.
Start Your QCode Plan