OpenAI GPT-5.4 mini and nano release notes
OpenAI has released GPT-5.4 mini and nano, two small-scale models designed to provide the reasoning and multimodal capabilities of the GPT-5.4 family for high-volume workloads. These models prioritize speed and cost-efficiency, making them suitable for applications where low latency is critical to the user experience.
GPT-5.4 mini: High-Performance Efficiency
GPT-5.4 mini is designed for professional tasks that require a balance of speed and intelligence. It runs more than 2x faster than GPT-5 mini and approaches the performance of the full GPT-5.4 model on several key evaluations, including OSWorld-Verified and SWE-Bench Pro.
Performance Benchmarks
GPT-5.4 mini (xhigh reasoning effort) demonstrates significant gains over its predecessor, GPT-5 mini (high reasoning effort):
- Coding: 54.4% on SWE-Bench Pro (Public) compared to 45.7% for GPT-5 mini.
- Computer Use: 72.1% on OSWorld-Verified, nearly matching the 75.0% achieved by GPT-5.4.
- General Intelligence: 88.0% on GPQA Diamond, compared to 81.6% for GPT-5 mini.
- Tool Use: 42.9% on Toolathlon and 93.4% on τ2-bench (telecom).
Multimodal and Computer Use Capabilities
GPT-5.4 mini is optimized for multimodal tasks, specifically the interpretation of dense user interfaces. It can quickly process screenshots to execute computer-use tasks, substantially outperforming GPT-5 mini on the OSWorld-Verified benchmark.
GPT-5.4 nano: Optimized for Low-Cost Tasks
GPT-5.4 nano is the smallest and most affordable version of the GPT-5.4 family. It is recommended for high-volume, simpler supporting tasks such as:
- Classification
- Data extraction
- Ranking
- Coding subagents handling simple tasks
On the GPQA Diamond benchmark, GPT-5.4 nano achieves 82.8%, surpassing the previous GPT-5 mini (81.6%).
Coding Workflows and Subagent Architecture
Both mini and nano models are optimized for fast iteration in coding workflows, including debugging loops, codebase navigation, and front-end generation.
The Subagent Pattern
OpenAI highlights a system architecture where larger models act as coordinators and smaller models act as executors. In this pattern, a model like GPT-5.4 handles high-level planning and final judgment, while delegating narrower subtasks—such as searching a codebase or reviewing large files—to GPT-5.4 mini subagents in parallel. This approach allows developers to compose systems that execute quickly at scale while maintaining high-level reasoning quality.
Availability and Pricing
GPT-5.4 mini and nano are available through various OpenAI platforms with the following specifications:
GPT-5.4 mini
- Availability: API, Codex, and ChatGPT.
- API Features: Supports text/image inputs, tool use, function calling, web search, file search, computer use, and skills. It features a 400k context window.
- API Pricing: $0.75 per 1M input tokens / $4.50 per 1M output tokens.
- Codex Integration: Available across the app, CLI, IDE extension, and web. It uses 30% of the GPT-5.4 quota, reducing costs for simpler coding tasks by approximately two-thirds.
- ChatGPT Integration: Available to Free and Go users via the ‘Thinking’ feature; serves as a rate limit fallback for GPT-5.4 Thinking for other users.
GPT-5.4 nano
- Availability: API only.
- API Pricing: $0.20 per 1M input tokens / $1.25 per 1M output tokens.