OpenAI Launches GPT-5.4 Mini and Nano Models
- •OpenAI releases GPT-5.4 mini and nano, high-efficiency models optimized for coding and subagents.
- •GPT-5.4 mini achieves 54.4% on SWE-Bench Pro while running twice as fast as GPT-5.4.
- •New nano model offers low-cost classification and data extraction at $0.20 per million input tokens.
OpenAI has expanded its flagship lineup with the debut of GPT-5.4 mini and nano, two lightweight models designed to deliver high-tier intelligence at a fraction of the cost and latency. While the primary GPT-5.4 model remains the powerhouse for complex reasoning, these smaller iterations target high-volume technical tasks. The mini version, in particular, bridges the gap between efficiency and capability, rivaling its larger sibling on specialized benchmarks like SWE-Bench Pro, which measures a model's ability to resolve real-world software issues.
The introduction of GPT-5.4 nano marks a new floor for cost-efficiency in the ecosystem, specifically tuned for narrow, repetitive tasks like text classification and ranking. This model is built to function as a "subagent," a specialized component within a larger AI system where a primary model handles the high-level planning while smaller models execute specific subtasks in parallel. By offloading these routine operations to a cheaper, faster model, developers can significantly reduce operational overhead without sacrificing overall system performance.
For users and developers, the shift toward "computer use" capabilities is equally significant. GPT-5.4 mini is capable of interpreting dense user interface screenshots, allowing it to navigate codebases and debug software with low latency. In practical terms, this means more responsive AI features in tools like Codex and ChatGPT, where the mini model is already being integrated to power the "Thinking" feature for free-tier users. This tiered approach suggests a future where AI utility is defined not just by raw power, but by the strategic deployment of specialized, efficient intelligence.