Introducing GPT-5.2-Codex

• Most advanced agentic coding model; targets professional software engineering.
  • State-of-the-art on SWE-Bench Pro (56.4%) and Terminal-Bench 2.0 (64.0%).
• Context compaction: preserves task state across extended multi-file sessions.
• Handles refactors, migrations, testing pipelines across large codebases.
• Strongest cybersecurity of any OpenAI model; vulnerability detection at codegen time.
  • Invite-only pilot for vetted defensive security professionals.
• Windows environment parity addressed; prior models were Unix-skewed.
• Pricing: $1.75/M input, $14/M output; 400K context, 128K output.
  • Available via ChatGPT Codex (paid) + API rollout; npm i -g @openai/codex.

X discourse

• @OpenAIDevs: “We’re retiring older models in Codex: gpt-5.2-codex, gpt-5.1-codex-mini, etc.” (4676 likes)
• @izutorishima: “GPT-5.3 Codex — コーディングスキルは人間を超えている.” (710 likes)
• @natolambert: “GPT 5.4 didn’t get enough praise for how big a step it was in OpenAI’s agent arc.” (298 likes)
• @haider1: “5.3 codex and gpt-5.2 xhigh were already ahead of opus 4.6 in analysis, agentic web search.” (275 likes)
• @Miles_Brundage: “GPT-5.4 made me less tolerant of Claude’s ‘it’s not X, it’s Y’ — another world is possible.” (223 likes)
• @haider1: “codex 5.3 and opus 4.6 pushed coding agents to a new level, then gpt-5.4 dropped.” (210 likes)

OpenAI (vb, OpenAI Developer Community — Jan 14, 2026) · ** · Read on openai.com