OpenAI has released GPT-5.4 mini across all Codex surfaces — the desktop app, CLI, IDE extension, and web interface. The model significantly improves over GPT-5 mini across coding, reasoning, image understanding, and tool use while running more than twice as fast.
The efficiency gains are the real story. GPT-5.4 mini uses only 30% of the rate limits that a full GPT-5.4 request consumes, making it ideal for rapid iteration, lightweight tasks, and subagent workflows where speed matters more than maximum capability. Despite the efficiency, it approaches GPT-5.4's performance on SWE-Bench Pro and OSWorld-Verified benchmarks.
Alongside the model release, OpenAI has shipped several Codex platform improvements. Plugin setup is smoother, with Codex now prompting to install missing plugins and honouring a configured suggestion allowlist. A new userpromptsubmit hook allows prompts to be intercepted, blocked, or augmented before execution — mirroring the hook system that Claude Code has offered. Custom prompts have been deprecated in favour of a skills system for reusable instructions and workflows, another convergence with patterns established by competing tools.
