On March 17, 2026, OpenAI officially released its latest small models, GPT-5.4 mini and GPT-5.4 nano.
Arriving just 12 days after the GPT-5.4 flagship release on March 5, this rollout came at a astonishing pace. Positioned by OpenAI as their most capable small models to date, these products represent an update that developers and enterprises cannot afford to overlook.
What is GPT-5.4 mini? — Near-flagship performance in a compact model
Over twice as fast as the prior generation, with performance on another level
The hallmark feature of GPT-5.4 mini is that while being more than twice as fast as the previous-generation GPT-5 mini, it achieves significant performance gains across coding, reasoning, multimodal understanding, and tool use.
Notably, it stands virtually shoulder to shoulder with the flagship GPT-5.4 model on select benchmarks. For a small model, this is an extraordinary achievement.
Capabilities seen through benchmarks
| Benchmark | GPT-5.4 (Flagship) | GPT-5.4 mini | GPT-5 mini (Previous generation) |
|---|---|---|---|
| SWE-Bench Pro (Coding) | 57.7% | 54.4% | 45.7% |
| OSWorld-Verified (PC operations) | 75.0% | 72.1% | 42.0% |
| Terminal-Bench 2.0 | — | 60.0% | — |
OSWorld-Verified is a demanding evaluation where "AI completes operations while inspecting PC screen screenshots," with a human baseline of 72.4%. GPT-5.4 mini's 72.1% virtually matches that.
Supported features
The API supports all of the following features:
- Text and image input (multimodal)
- Tool use and function calling
- Web search and file search
- Computer Use (PC operation automation)
- Skills (granting custom capabilities)
What is GPT-5.4 nano? — Maximizing cost and speed efficiency
The smallest, most affordable, and fastest in the GPT-5.4 family
GPT-5.4 nano is the most compact model in the GPT-5.4 family. Designed for tasks prioritizing speed and cost above all else, it is well-suited for classification, data extraction, ranking, and simple coding subagents.
nano benchmarks
| Benchmark | GPT-5.4 nano | GPT-5 mini (Previous generation) |
|---|---|---|
| SWE-Bench Pro | 52.4% | 45.7% |
| Terminal-Bench 2.0 | 46.3% | — |
| OSWorld-Verified | 39.0% | 42.0% |
While it trails the prior generation on OSWorld-Verified (PC operations), it surpasses it on coding benchmarks. Rather than PC operations, nano is designed for high-speed, high-volume workloads such as text processing and code analysis.
Additionally, there are reports that when run at maximum reasoning effort, GPT-5.4 nano outperforms the previous-generation GPT-5 mini.
Price comparison — Remarkably cost-competitive
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| GPT-5.4 mini | $0.75 | $4.50 |
| GPT-5.4 nano | $0.20 | $1.25 |
GPT-5.4 nano's price of $0.20 per million input tokens is cheaper than Google's Gemini 3.1 Flash-Lite. By calculation, approximately 76,000 photos could be described for $52.
Where can it be used? — Available platforms
GPT-5.4 mini
- ChatGPT (as the "Thinking" feature for Free and Go users)
- Codex (more economical to use, consuming only 30% of the GPT-5.4 quota)
- OpenAI API
GPT-5.4 nano
- OpenAI API only (unsupported in the ChatGPT UI)
The "subagent" pattern enters practical viability
A highlight of this release is bringing subagent architectures pairing large and small models into practical reality.
Concrete architecture example (Codex)
GPT-5.4（旗艦）
└── タスク計画・調整・最終判断
GPT-5.4 mini（サブエージェント × 複数）
├── コードベースの検索
├── 大きなファイルのレビュー
└── 補足ドキュメントの処理Because using GPT-5.4 mini in Codex consumes only 30% of the GPT-5.4 quota, delegating simpler subtasks to mini significantly lowers costs while boosting productivity.
Caveat: mini is not a complete replacement for GPT-5.4
mini cannot replace the flagship model across every task.
In ultra-long context tracking accuracy (OpenAI MRCR v2 / 8 needles / 64K–128K):
| Model | Score |
|---|---|
| GPT-5.4 (Flagship) | 86.0% |
| GPT-5.4 mini | 47.7% |
For tasks requiring simultaneous tracking of numerous details across very long documents, using the flagship model continues to be recommended.
Summary: The "new normal" for small models
The release of GPT-5.4 mini and nano brings tangible changes to how AI is utilized.
| Key point | Details |
|---|---|
| Cost reduction | nano is more affordable than Gemini Flash-Lite |
| Speed improvement | mini is more than 2x faster than prior generation |
| Performance enhancement | Significantly outperforms previous generations in coding and reasoning |
| Practical configuration | Large + small subagent design reaches practical viability |
The basic guideline is to use nano if you want to process high volumes as cheaply as possible, and mini if you want a fast, smart model while keeping costs down.
Going forward, as small models further improve in performance and gain speed, architectures where large models make decisions and small models execute at scale will increasingly become mainstream.
At GleamHub Inc., we provide system development and consulting utilizing the latest AI models like these. Please feel free to contact us for any inquiries regarding AI adoption.
