OpenAI has introduced two lightweight models, GPT-5.4 Mini and GPT-5.4 Nano, positioning both as execution-layer components inside hybrid AI systems rather than flagship models. The company describes the approach as trading some precision for major gains in speed and cost efficiency, a design aimed at multi-agent setups where stronger models handle planning and smaller ones carry out large volumes of routine work in parallel.
GPT-5.4 Mini targets broad execution tasks
GPT-5.4 Mini is the central release in this launch. OpenAI says the model delivers 2x faster reasoning than its predecessor and improves code generation, multimodal understanding, and tool use. In benchmark results, Mini scored 54.4% on SWE-Bench Pro and 72.1% on OSWorld-Verified.
Its feature set includes mixed text-and-image input, tool calling, web and file search, and computer UI interaction. OpenAI lists code assistants, IDE integrations, code review, screenshot interpretation, computer-use sub-agents, and real-time interactive systems among the intended use cases. Pricing is set at $0.75 per million input tokens and $4.50 per million output tokens. The model is available now across ChatGPT free and Go tiers, Codex, and the OpenAI API, with deployment also supported through Azure AI Foundry.
GPT-5.4 Nano goes after low-latency, high-throughput workloads
GPT-5.4 Nano has a narrower role. OpenAI calls it the company’s smallest and most cost-efficient model, built for use cases that are highly latency-sensitive and require large-scale parallel processing. On benchmarks, Nano posted 52.4% on SWE-Bench Pro and 39.0% on OSWorld. Its scores trail Mini, but its price point is much lower.
Nano supports instruction following, function calling, basic coding, image understanding, classification, and data extraction. OpenAI says it is suited to structured workloads, not complex desktop operations or deeper reasoning tasks. The model is priced at $0.20 per million input tokens and $1.25 per million output tokens, roughly one quarter of Mini’s pricing. The company points to classification and data extraction, coding sub-agent support, high-volume automation, request routing, form processing, and customer service workflows as key scenarios. Nano is currently available only through the API and does not appear in the ChatGPT interface.
Built around the sub-agent model architecture
OpenAI framed both releases as products for the “sub-agent era.” In that structure, flagship systems such as GPT-5 or the o-series reasoning models handle planning and higher-level judgment, while repetitive and structured subtasks like web extraction, data transformation, form filling, and code snippet generation are assigned to smaller, faster, and cheaper models. The emphasis on tool calling and computer interaction in GPT-5.4 Mini and Nano reflects that system design.