OpenAI drops GPT-5.4 mini and nano

OpenAI is turning the mini tier into the execution engine for agentic software, not just a cheap fallback. The real shift is that builders can now reserve the big model for planning and let mini do the fast, repetitive work.

• –OpenAI says GPT-5.4 mini is more than 2x faster than GPT-5 mini while improving across coding, reasoning, multimodal understanding, and tool use.
• –It reaches 54.4% on SWE-Bench Pro and 72.1% on OSWorld-Verified, which is strong enough to matter in real coding and computer-use loops.
• –In the API it costs $0.75 per 1M input tokens and $4.50 per 1M output tokens, and in Codex it burns only 30% of GPT-5.4 quota.
• –The subagent angle is the real product strategy: use mini for search, review, and support tasks so the top model stays focused on planning and judgment.