OpenAI Codex & GPT-5.2-Codex / Sora Android

OpenAI released GPT-5.2-Codex as an Agentic Coding Model optimized for autonomous engineering tasks. - Focus: complex long-term tasks, security, visual UI understanding - Context compression enables long-context projects - Stability for iterative execution preserved - Enhanced security tooling for vulnerability discovery - Visual parsing of screenshots and UI to prototype code - Benchmarks: SWE-Bench Pro 56.4%, Terminal-Bench 64.0% - Simon Willison used Codex CLI + GPT-5.2 to port JustHTML to JS with ~9000 tests passing - Simon’s process used 140M tokens and 43 commits in closed loop - OpenAI built Sora Android in 28 days with Codex; ~85% of code generated by AI - Sora first-day: 1M+ videos generated, 99.9% crash-free - Teams used architecture-first, spec-first, automated tests and CI loops
Key insights
GPT-5.2-Codex positioned as an agentic coding model
OpenAI announced GPT-5.2-Codex, framed as an Agentic Coding Model for autonomous engineering tasks with three highlighted improvements: handling complex long-running tasks via context compression, enhanced defensive network-security assistance, and improved visual/interactive UI understanding for converting designs into prototypes. Published benchmarks include SWE-Bench Pro at 56.4% and Terminal-Bench 2.0 at 64.0%, and OpenAI noted Windows native operation improvements.
Real-world porting case shows agentic coding strengths
Simon Willison used Codex CLI and GPT-5.2 to port JustHTML from Python to JavaScript, running 9,200+ html5lib-tests and producing ~9,000 lines of code across 43 commits with about eight prompts. The workflow leveraged test suites, spec-first planning, smoke tests and CI loops so the agent could iterate against failing tests until green.
Sora Android demonstrates rapid productization with Codex
OpenAI described building Sora for Android in 28 days with a four-person team where approximately 85% of the code was AI-generated. The team supplied architectural scaffolding, internal 'agent manuals' and CI-based test loops, resulting in 1M+ videos generated on day one and a 99.9% crash-free rate. The process emphasized planning, architecting first, and treating agents as junior engineers to manage quality.
Agentic workflows require architecture and test automation
Across examples, teams adopted 'spec-first' and architecture-first approaches: provide AI with system design, smoke tests, CI-triggered test loops, and per-session artifacts to maintain continuity. This pattern reduces agent errors, enables automated verification, and shifts human roles toward oversight, planning, and integration rather than manual coding.