GPT-5.4 nano
Cheapest GPT-5.4-class model for simple high-volume tasks such as extraction, ranking, and lightweight subagents.
You always get the exact model you pick — we never silently route you to another.
About GPT-5.4 nano
At roughly $0.18 per million tokens blended, GPT-5.4 nano occupies a deliberate efficiency tier in OpenAI's GPT-5.4 family — not a stripped-down mini, but a purpose-built layer for systems that need speed and volume rather than depth. It runs at 146.5 tokens per second, well above the peer median, and its MATH-500 score of 88.6% makes it surprisingly capable for calculation-heavy pipelines. Developers building classification engines, extraction jobs, and routing layers consistently cite cost and throughput as the main reasons to reach for nano over its larger siblings. That said, it is not a generalist. Users have found real gaps: XML and LaTeX structured output degrades sharply compared to JSON, spatial reasoning tasks expose clear limits, and its 6-second time-to-first-token is notably higher than comparable models. For agentic workflows requiring complex image interpretation or multi-step reasoning, the larger GPT-5.4 mini will serve better. Nano earns its place as a high-volume component in layered systems, not a sole workhorse.
Best for
- Data extraction and document parsing at scale, where per-token cost and throughput matter more than nuanced comprehension
- Text classification and routing — email triage, ticket categorization, content moderation, and query labeling in high-frequency pipelines
- JSON-structured output workflows, including form extraction and API-feeding tasks that depend on reliable structured generation
- Lightweight coding subagents within larger orchestration systems where nano handles simple generation and refactoring passes cheaply
- Math and calculation-heavy tasks where its 88.6% MATH-500 performance adds genuine value without the cost of a larger model
Specifications
| Provider | OpenAI |
|---|---|
| Released | 2026-03 |
| Intelligence | Medium |
| Speed | Medium |
| Context window | 400,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | August 31, 2025 |
| Input price | $0.20 / 1M tokens |
| Output price | $1.25 / 1M tokens |
| Request cost | 2 base requests |
| Plan tier | Base |
| Input | Text, Image |
| Output | Text |
| Features | Cached input: $0.02 / 1M tokens, Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, and MCP supported, Computer use and tool search are not supported, Fine-tuning not supported; distillation supported |
| API endpoints | v1/chat/completions, v1/responses, v1/realtime, v1/assistants, v1/batch |
| Model ID | gpt-5.4-nano |
Frequently asked questions
Input is $0.20 per million tokens and output is $1.25 per million tokens, putting the blended effective cost at around $0.18 per million tokens — the cheapest option in the GPT-5.4 family.
400,000 tokens input with a maximum of 128,000 tokens output.
High-volume classification, data extraction, JSON-structured output, ranking, and serving as a fast, cost-efficient subagent layer inside larger systems. It also holds up well on mathematical tasks.
Spatial and 3D reasoning, complex agentic computer-use tasks (39% on OSWorld-Verified), XML and LaTeX structured output, and competitive math olympiad problems like AIME where scores sit at 26–30%.
Nano is cheaper and faster in tokens-per-second throughput, but mini handles complex generation, deeper image understanding, and harder reasoning tasks more reliably. Nano is designed as an efficiency tier within a multi-model system, not a replacement for mini.
No. It is API-only and is not accessible through ChatGPT consumer products. Using it requires direct developer integration.