What happened
On March 22, 2026, OpenAI released GPT-5.2-Codex as a specialized coding model built for long-horizon agentic programming tasks. The model achieves state-of-the-art performance on SWE-Bench Pro and Terminal-Bench 2.0 — benchmarks designed to test realistic terminal-based agentic performance. Key improvements over previous Codex models include stronger context compaction for extended sessions, more reliable completion of large refactors and code migrations, improved vision capabilities for interpreting screenshots and technical diagrams, and significantly stronger cybersecurity capabilities. The model is available through the Codex product and the OpenAI API.
Why it matters
GPT-5.2-Codex marks a continuation of OpenAI's dedicated coding model track — a signal that the company views agentic software development as a distinct problem space worth specializing for. The long-horizon improvements (tracking state across large refactors without losing context) address one of the most common failure modes in AI coding agents: losing the thread when plans shift mid-execution. The cybersecurity focus is notable for enterprise teams that need coding agents to work in security-sensitive contexts. Combined with the simultaneous Astral acquisition, this suggests OpenAI is assembling a vertically integrated coding stack.
Who should pay attention
- Engineering teams using Codex or the OpenAI API for automated code review, migration, and refactoring tasks
- Security engineers evaluating AI-assisted vulnerability research and code hardening workflows
- Developers building coding agents who need a model with strong context compaction for multi-step tasks
- Anyone benchmarking AI coding models for production deployment decisions