GPT-5.3-Codex — OpenAI’s First High-Cybersecurity Model, Same Day as Opus 4.6
OpenAI released GPT-5.3-Codex (Feb 5): ~25% faster; vendor SOTA on SWE-Bench Pro and Terminal-Bench 2.0 (77.3%); multi-day autonomous app builds with real-time steering. First OpenAI model classified High for cybersecurity under its Preparedness Framework with expanded cyber stack. Agentic coding as core infra—not autocomplete.
Correction (October 14, 2026): Specified that the High designation concerns cybersecurity and added the vendor benchmark reasoning-effort setting.
Two labs shipped major agentic coding upgrades the same day. OpenAI’s proprietary stake: coding agents as core software-team infrastructure, with OpenAI’s first High cybersecurity-capability Preparedness designation—capability and cyber controls announced together.
OpenAI introduced GPT-5.3-Codex on February 5. Pitch: unify Codex-line coding advances with GPT-5.2-class professional reasoning; ~25% faster (company). All release evaluations used xhigh reasoning effort. Vendor benches: leads public SWE-Bench Pro; Terminal-Bench 2.0 77.3%; strong OSWorld-Verified computer use; matches GPT-5.2 on GDPval knowledge-work tasks. Long-horizon: iterate on full apps/games over days and millions of tokens; skills for web dev, debugging, creative generation; real-time steering and progress updates in updated Codex app. Internal dogfood: early versions used to debug training runs, optimize infra, analyze evals, help build the release. Preparedness: first OpenAI model classified High capability for cybersecurity; expanded cyber safety stack (safety training, monitoring, trusted access pilots for defensive research, threat-intel enforcement); Trusted Access for Cyber pilot launched alongside.
Claims vs checks#
Product features and High designation are OpenAI primary. SWE-Bench Pro / Terminal-Bench / OSWorld / GDPval numbers are vendor-cited; independent public-board results are not established here. “25% faster” is company relative claim (baseline unspecified in brief). Dual-use cyber framing is OpenAI’s risk taxonomy, not a regulator’s.
Limits#
- Multi-day autonomy demos are curated; production reliability differs.
- High preparedness ≠ public exploit capability scores.
- Same-day Opus 4.6 comparisons require matched scaffolds—don’t mash leaderboards casually.
Sources#
- Anthropic release notes.
- OpenAI: “Introducing GPT-5.3-Codex” (February 5, 2026). Primary.