OpenAI launches GPT-5.4 mini and nano for coding, subagents and high-volume AI workloads

OpenAI has released GPT-5.4 mini and GPT-5.4 nano, two compact versions of its flagship lineup built not for record-setting reasoning but for fast, high-volume workloads. The company is betting on models that are cheaper and faster while keeping a sizable share of the full-size GPT-5.4's capabilities.

What mini does well

GPT-5.4 mini is the larger of the two new models. According to OpenAI, it improves markedly over GPT-5 mini in programming, reasoning, tool use and multimodal analysis, and it is more than twice as fast. On the SWE-Bench Pro benchmark the model scored 54.4%, versus 45.7% for GPT-5 mini. On OSWorld-Verified, which tests the ability to work with interfaces from screenshots, it scored 72.1%, versus 42.0% for the previous mini version.

This matters beyond the charts. OpenAI explicitly positions GPT-5.4 mini as a working model for tasks where users feel latency: code autocompletion and edits, fast debugging loops, subagents for auxiliary operations, and systems that read screenshots and interact with an interface. The idea is simple: not every task needs the most expensive model if a smaller one handles it almost as well, only noticeably faster.

Where nano fits

GPT-5.4 nano is the smallest and cheapest model in the new lineup. OpenAI recommends it not as a general-purpose chat engine but as a utility tool for simple but frequent operations. These are scenarios where throughput matters more than depth of reasoning: processing streams of documents, classification, field extraction, ranking results and supporting simple coding subtasks. Even nano scored 52.4% on SWE-Bench Pro, also clearly ahead of GPT-5 mini.

Together, mini and nano fit an architecture in which one large model plans the work and several small ones carry it out in parallel. That is the scenario OpenAI is promoting in Codex: GPT-5.4 can coordinate the process while GPT-5.4 mini takes on narrow tasks such as searching a codebase, reading large files and processing documentation. This split approach helps keep both latency and budget under control.

• Fast AI assistants for writing and editing code
• Subagents for repository search and analysis of large files
• Tools that understand screenshots and operate interfaces
• Classification, data extraction and ranking in high-volume pipelines
• Cheaper execution of auxiliary tasks without lowering overall system quality

Access and pricing

As of March 17, 2026, GPT-5.4 mini is available in the API, Codex and ChatGPT from day one. In the API the model supports text, images, function calling, web search, file search, computer use and skills. The context window is 400 thousand tokens, and pricing is $0.75 per million input tokens and $4.50 per million output tokens. For services handling many parallel requests, this is one of the main selling points.

In Codex, mini uses only 30% of the quota of the full-size GPT-5.4, so it can be used for cheap parallel subtasks. In ChatGPT the model has a more limited role: for Free and Go users it is available through Thinking mode, and for everyone else it serves as a fallback when the GPT-5.4 Thinking limit is reached.

GPT-5.4 nano, for its part, is available only through the API and costs $0.20 per million input tokens and $1.25 per million output tokens.

What it means

OpenAI is increasingly building its lineup not around a single "best" model but around a set of roles: the big model thinks and coordinates, the small ones quickly handle the routine. For developers and AI products, that is good news: the cost of agentic systems can be cut without a sharp drop in quality, especially where speed, parallelism and high request volume matter.