OpenAI just dropped GPT-5.3-Codex, their most capable agentic coding model yet. It blends GPT-5.2-Codex’s top-tier coding chops with GPT-5.2’s reasoning and professional knowledge. Think of it as a colleague you can steer mid-task without losing context—great for long-running jobs involving research, tool use, and complex execution. It’s classified as ‘High capability’ in biology (like other recent models) and, for the first time, in cybersecurity under their Preparedness Framework. OpenAI says they don’t have definitive proof it hits the high threshold, but they’re being cautious, activating layered safeguards to block threat actors while making the tech easy for cyber defenders. The model doesn’t reach high capability for AI self-improvement. Why it matters: This proactive safety-first launch signals a shift where even potential risks trigger protections, balancing powerful AI deployment with responsible guardrails in critical domains.