OpenAI GPT-5.2-Codex and coding agents

OpenAI released GPT-5.2-Codex on Dec 18, 2025 targeting agentic coding with long-run task stability, security features and visual UI understanding.
Key insights
GPT-5.2-Codex positioned as agentic coding model
OpenAI announced GPT-5.2-Codex as an Agentic Coding Model focused on complex long-running engineering tasks, network security assistance, and improved visual and UI understanding. The model includes native long-context compression to maintain continuity during large refactors, iterative stability to persist through plan changes, and added deployment safety measures. OpenAI reported state-of-the-art results on coding agent benchmarks.
Benchmark performance and stated metrics
OpenAI reported GPT-5.2-Codex achieves 56.4% on SWE-Bench Pro and 64.0% on Terminal-Bench 2.0, with notable improvements over GPT-5.1 in agentic coding tasks and enhanced Windows native operation.
Real-world coding agent case: Simon Willison uses Codex CLI + GPT-5.2
Simon Willison used Codex CLI together with GPT-5.2 to port JustHTML from Python to JavaScript, running over 9,200 html5lib-tests. The process produced ~9,000 lines of code across 43 commits, driven by about eight prompts. The workflow relied on test-driven loops: AI writes code, runs tests, reads failures, and iterates until green, demonstrating the coding agent 'comfort zone' of translation with comprehensive test suites and existing architecture.
Coding agent practical guidance
Practical guidance for coding agents includes: prefer translation or 'fill-in' tasks with existing reference implementations, provide automated verification tools (tests, linters, runtimes), design architecture to split tasks into context-sized chunks, and use spec-first and smoke-test loops to guide iterative agent development.