OpenAI drops GPT-5.4 mini and nano for high-speed coding

If you build with AI agents or coding assistants, you already know the painful tradeoff: the smartest models are too slow and expensive, while the cheap ones make silly mistakes.

OpenAI just launched GPT-5.4 mini and GPT-5.4 nano to close that exact gap.

These are lightweight models built specifically for latency-sensitive tasks like coding copilots, subagents, and real-time computer use. And according to OpenAI's numbers, the performance cut is surprisingly small.

Small size, flagship numbers

GPT-5.4 mini runs more than two times faster than its predecessor, GPT-5 mini. Yet on multiple tough benchmarks, it breathes down the neck of the flagship GPT-5.4.

On the SWE-Bench Pro coding benchmark, GPT-5.4 scores 57.7%, while the new mini hits 54.4%. For comparison, the older GPT-5 mini sits at 45.7%.

Even more striking is computer control: on OSWorld-Verified, which tests taking screenshots and operating dense software interfaces, GPT-5.4 mini hits 72.1%, just trailing the full model's 75.0%.

Then there is GPT-5.4 nano, OpenAI's cheapest offering yet. It trades off deep reasoning for raw speed, targeting classifications, extraction, and lightweight coding subtasks.

Both models feature a massive 400k token context window. In terms of API pricing, mini costs $0.75 per million input tokens and $4.50 for output. Nano drops that down to $0.20 input and $1.25 output.

Mini is available immediately across the API, ChatGPT, and Codex. In Codex, it burns just 30% of your GPT-5.4 quota. Nano is available today in the API only.

Why it matters

We think this release signals a huge shift in how modern AI architectures will be built.

Until now, running multi-agent workflows was often prohibitively expensive or painfully laggy. Having a flagship model do everything from coordinating tasks to reading tiny config files burns cash and patience fast.

With GPT-5.4 mini, the "hub and spoke" model finally makes economic sense. You keep a frontier model like GPT-5.4 in the driver's seat for overarching logic, but delegate repetitive subtasks—like searching files, testing code snippets, or parsing visual interfaces—to a squad of mini agents running in parallel.

For developers and startups, it is definitely worth switching high-frequency background pipelines over immediately. When a 75-cent model delivers over 90% of a flagship's coding and GUI capability at twice the speed, running frontier models on brute-force tasks just feels wasteful.