OpenAI has officially launched GPT-5.4 mini and GPT-5.4 nano, releasing its most capable small models designed to handle high-volume, latency-sensitive workloads.
The new mini iteration offers a significant performance upgrade over the previous GPT-5 mini across reasoning, coding, tool use, and multimodal understanding, while running more than twice as fast.
These models are engineered for applications where speed directly shapes the user experience, such as responsive coding assistants, real-time multimodal applications, and systems that capture and interpret screenshots.
The release proves that the largest model is not always the best choice; instead, systems that respond quickly and use tools reliably are often superior for complex professional workflows.
Both models demonstrate exceptional effectiveness in coding environments that require rapid iteration, including codebase navigation, debugging loops, generating front-end code, and targeted edits.
Benchmarks show that GPT-5.4 mini approaches the accuracy of the flagship GPT-5.4 model on evaluations like SWE-Bench Pro, delivering one of the strongest performance-per-latency tradeoffs for developers.
A major technical highlight is their integration into subagent architectures. On platforms like Codex, developers can deploy a larger model like GPT-5.4 for complex planning, coordination, and final judgment, while delegating narrower tasks to GPT-5.4 mini-subagents.
These smaller agents can process supporting documents, search codebases, and review large files in parallel.
This composition allows systems to execute faster and scale efficiently without relying on a single massive model for minor operations.
GPT-5.4 mini brings substantial improvements to multimodal tasks, specifically in computer-use scenarios. The model can rapidly analyze dense user interface screenshots to execute actions with high precision.
On the OSWorld-Verified benchmark, GPT-5.4 mini achieved an accuracy of 72.1 percent, nearly matching the 75.0 percent score of the larger GPT-5.4 and vastly outperforming the 42.0 percent achieved by the older GPT-5 mini.
For simpler support tasks, GPT-5.4 nano is the smallest and most cost-effective option. OpenAI recommends this model for data extraction, classification, ranking, and lightweight coding tasks where speed and cost efficiency are critical.
GPT-5.4 mini is currently available through the OpenAI API, Codex, and ChatGPT. Within the API, it features a massive 400k context window, supporting text and image inputs, function calling, web search, and computer use.
The cost is set at 0.75 dollars per one million input tokens and 4.50 dollars per one million output tokens.
In Codex, developers can handle routine coding tasks using GPT-5.4 mini for about one-third the cost, utilizing only 30 percent of the standard GPT-5.4 quota.
ChatGPT Free and Go users can access the model via the Thinking feature. At the same time, it serves as a rate limit fallback for other tiers.
The nano variant remains available exclusively via the API, priced at 0.20 dollars per one million input tokens and 1.25 dollars per one million output tokens. This formatting aligns with your preference for professional, table-free technical news structures.
Follow us on Google News, LinkedIn, and X for daily cybersecurity updates. Contact us to feature your stories.
