OpenAI launched its two most powerful small models to date, GPT-5.4 mini and GPT-5.4 nano, on Tuesday, significantly narrowing the performance gap with flagship models at lower latency and cost.GPT-5.4 mini surpasses the previous generation GPT-5 mini across core dimensions such as programming, reasoning, multimodal understanding, and tool invocation, with a speed increase of over 2 times, and approaches the performance of the larger GPT-5.4 in benchmark tests like SWE-Bench Pro.GPT-5.4 nano is positioned as the lowest-cost, lowest-latency lightweight option, available to developers solely through API, designed specifically for data classification, extraction, and simple programming sub-tasks.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/45a3c9cd-272f-4332-a2d0-5849b8b57afe.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/45a3c9cd-272f-4332-a2d0-5849b8b57afe.png"/>The launch of these two models aims to fill the gap in real-time interaction scenarios where large models struggle to be implemented due to high latency, directly impacting the rapidly growing commercial markets covering programming assistants, AI agent systems, and multimodal applications.<h2>mini for consumer end, nano exclusive to API</h2>GPT-5.4 mini is now available across three channels: OpenAI API, Codex platform, and ChatGPT.The API pricing for GPT-5.4 mini is $0.75 per million input tokens and $4.50 per million output tokens, supporting text and image input, tool invocation, function calls, web searches, file retrieval, computer control, and skill expansion, with a context window of up to 400,000 tokens.On the Codex platform, GPT-5.4 mini consumes only 30% of the GPT-5.4 quota, reducing the cost for developers handling simple programming tasks to about one-third of the flagship model. Codex also supports delegating workloads to sub-agents running on GPT-5.4 mini, allowing tasks with lower reasoning density to automatically fall to the cheaper model.On the ChatGPT side, Free and Go users can select the &#34;Thinking&#34; feature to use GPT-5.4 mini through the &#34;+&#34; menu; other paid users will have this model activated as an automatic downgrade option once the GPT-5.4 Thinking throughput rate limit is reached.GPT-5.4 nano is currently available only through API for developers, priced at $0.20 per million input tokens and $1.25 per million output tokens, making it the lowest-priced among the two new models. OpenAI states that nano is suitable for scenarios where higher-level models coordinate and manage sub-agents responsible for handling secondary support tasks.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/6026472f-156f-40da-b9b3-50ce74c74435.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/6026472f-156f-40da-b9b3-50ce74c74435.png"/><h2>mini approaches flagship, nano surpasses previous generation</h2>From the evaluation data released by OpenAI, the performance of GPT-5.4 mini in programming and multimodal tasks is particularly outstanding.In the programming benchmark SWE-bench Pro, mini scored 54.4%, narrowing the gap with GPT-5.4's 57.7% to 3.3 percentage points, significantly higher than GPT-5 mini's 45.7%.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c7fdf84b-fed0-4019-b07a-e616678bae26.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c7fdf84b-fed0-4019-b07a-e616678bae26.png"/>In the computer control benchmark OSWorld-Verified, mini approached GPT-5.4's 75.0% with a score of 72.1%, significantly ahead of GPT-5 mini's 42.0%.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d90f4559-26c5-4d56-8f14-0e467ba915d6.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d90f4559-26c5-4d56-8f14-0e467ba915d6.png"/>In terms of tool invocation capability, GPT-5.4 mini scored 93.4% in the τ2-bench telecommunications test, a significant improvement over GPT-5 mini's 74.1%. In the general intelligence test GPQA Diamond, mini scored 88.0%, while nano also reached 82.8%, both surpassing GPT-5 mini's 81.6%.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b0dc1040-eb5e-4dd1-ba60-3628e17b31af.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b0dc1040-eb5e-4dd1-ba60-3628e17b31af.png"/>It is noteworthy that GPT-5.4 nano performed worse than GPT-5 mini in some visual tasks, with an OSWorld-Verified score of 39.0%, lower than the latter's 42.0%. However, in programming and tool invocation tasks, nano still achieved significant improvements over the previous generation.<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d9cb7102-70b4-48fc-b191-2a3d3afae583.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d9cb7102-70b4-48fc-b191-2a3d3afae583.png"/>OpenAI stated that the design priority of nano is low latency and low cost, rather than overall performance, and developers need to weigh trade-offs based on specific tasks when making selections.<h2>Sub-agent architecture, multi-model collaboration as a new paradigm for product design</h2>OpenAI emphasized the position of the two new models in a multi-model hierarchical system in the release materials.Taking its self-developed programming assistant Codex as an example, GPT-5.4 is responsible for planning, coordinating, and final judgment, while the GPT-5.4 mini sub-agent processes finer-grained sub-tasks such as codebase retrieval, large file review, and document assistance in parallel.OpenAI stated that with smaller models being faster and more powerful, developers no longer need to use a single model to handle all tasks, but can build systems where large models are responsible for decision-making while smaller models execute tasks quickly and at scale. OpenAI said:<blockquote>GPT-5.4 mini is the most powerful small model we have developed for this workflow to date.</blockquote>This architecture is particularly critical for high-concurrency work, where response latency directly affects product experience in scenarios such as programming assistance, screenshot analysis, and real-time image understanding. The optimal choice is often not the most capable model, but rather the one that achieves the best balance between speed, tool reliability, and task performance.For developers, the release of GPT-5.4 mini and nano means a clearer path to significantly reducing inference costs without sacrificing the overall intelligence level of the system

OpenAI

OpenAI stated that as smaller models become faster and more powerful, developers no longer need to use a single model to handle all tasks, but can build systems where large models are responsible for decision-making, while smaller models execute tasks quickly and at scale. "GPT-5.4 mini is the most powerful small model we have developed for this workflow so far."

OpenAI releases GPT-5.4 mini and nano, approaching flagship model performance at a lower cost