We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Mar 17·5 min read
Introducing GPT-5.4 mini and nano
TL;DR
Large language models are computationally expensive to run at scale, creating friction for high-volume applications and cost-sensitive deployments. OpenAI released GPT-5.4 mini and nano—smaller, optimized variants of GPT-5.4 designed for coding, tool use, multimodal reasoning, and sub-agent workloads.
✦ Why It Matters
Use GPT-5.4 mini or nano to reduce API costs and latency for coding, tool use, and agentic workloads without sacrificing capability.
Key Takeaways
- Deploy GPT-5.4 mini or nano for coding and tool-use tasks to cut latency and API costs.
- Use nano for high-volume, cost-sensitive sub-agent or delegated reasoning workflows.
- Evaluate mini and nano on your specific use case before migrating from full GPT-5.4.
- Monitor token usage and latency metrics post-deployment to validate cost and performance gains.
Full Summary
Large language models (LLMs—neural networks trained on vast text data to predict and generate language) have become central to production systems, but their computational cost limits deployment at scale. OpenAI introduced GPT-5.4 mini and nano, distilled versions of the full GPT-5.4 model optimized for specific workloads: coding tasks, tool integration (calling external APIs or functions), multimodal reasoning (processing text, images, and other data types together), and sub-agent architectures (systems where smaller models handle delegated tasks).
These models achieve faster inference (response time) and lower per-token costs compared to the full model. The optimization targets high-volume API usage and agentic workflows where latency and cost directly impact user experience and operational budgets.
Engineers can now deploy capable models for specialized tasks without the overhead of running full-scale variants.