GPT-5.3-Codex — What’s New
Released by OpenAI on 2026-02-05 — the same day as Claude Opus 4.6.
Official links
- Introducing GPT-5.3-Codex: https://openai.com/index/introducing-gpt-5-3-codex/
- System card: https://openai.com/index/gpt-5-3-codex-system-card/
Key updates
A unified model
GPT-5.3-Codex combines Codex-grade coding with GPT-style general reasoning. The goal: fewer model switches depending on the task.
25% faster (reported)
OpenAI reports a ~25% speedup for Codex users. In agent workflows, small per-turn latency savings add up quickly.
“Self-involved” development
OpenAI describes this as the first model that participated in parts of its own creation pipeline—debugging training processes, deployment operations, diagnostics, and evaluations.
Real-time interactive coding
More interactive execution: the model can frequently report progress and decisions, and developers can interrupt/redirect during the run (more like collaborating with a teammate).
Cybersecurity positioning
OpenAI notes elevated safeguards under the Preparedness Framework and emphasizes cautious language around thresholds; the posture suggests concern about rapidly increasing capability.
Long-running task execution
Better support for research + tool use + complex multi-step execution chains.
Availability
| Channel | Status |
|---|---|
| ChatGPT paid plans (app / CLI / IDE extension / web) | Available |
| API | “Rolling out with safety gating” (per OpenAI) |
What this means for me
A unified model is the most practical change: fewer decisions about “which model do I use for this?”
The speed gain is underrated. Agent tasks are multi-turn; shaving seconds per turn can mean minutes per task.
Real-time interaction also changes the workflow: instead of “send a spec and wait,” it’s closer to live collaboration—better for ambiguous requirements.
API availability lag is the main downside for automation setups. For now, the realistic plan is: use what’s available today, and switch once the API path is open.