GPT-5.4 mini and nano: Benchmarks, Access, and Reactions

The recent news from OpenAI is especially interesting for users of GPT-5 mini: the new GPT-5.4 mini model is twice as fast and brings an improved performance across all areas. Additionally, OpenAI released the newest version of their most efficient model class, GPT-5.4 nano.

In this article, we will walk you through what GPT-5.4 mini and nano are, how they perform compared to GPT-5.4, and who can profit from OpenAI’s newest “fast lane” models.

Make sure to also check out our comparison of GPT-5.4 vs Claude Opus and our guide to OpenAI Frontier.

What Is GPT-5.4 mini?

GPT-5.4 mini is OpenAI’s newest small LLM, replacing GPT-5 mini. It brings several significant improvements to the performance of its predecessor, while being twice as fast, which is one of the main selling points.

In the API, it supports a wide range of features: 

• Text and image inputs
• Tool use and function calling
• Web search
• Computer use
• Skills

What Is GPT-5.4 nano? 

GPT-5.4 nano is the smallest version of OpenAI’s newest model line, and it replaces GPT-5 nano. As the name might give away, it is even more efficient than the mini model, at a lower performance. Still, the GPT-5.4 nano beats the old mini model, GPT-5 mini, in many benchmarks.

It doesn’t support as many features as the mini model, but it offers today’s standard API features like image input, tool use, function calling, and structured outputs.

Who Is GPT-5.4 mini and nano Really For?

The new release offers some flexibility in choosing the right model, considering the classic tradeoff between performance on the one hand, and latency and price on the other.

OpenAI recommends mini and nano for developers who are working on applications where you don’t want lag. Basically, things that need to feel responsive, where users are going to be really intolerant of delay. 

For reasoning-heavy tasks with little room for error, multimodality, and agentic tasks, GPT-5.4 remains the first choice. 

GPT-5.4 mini and nano Benchmarks

Let’s take a look at the LLM benchmarks. Some notable results:

• Coding: Both 5.4 mini (54.4%) and nano (52.4%) reach a SWE‑Bench Pro score of over 50% and don’t lag behind GPT-5.4 by a lot. The improvement against GPT-5 mini (45.7%) is significant.
• Terminal agents: In Terminal‑Bench 2.0, the distance between the three different flavors of 5.4 models really shows. GPT-5.4 mini (60.0%) can compete with previous flagship models, such as GPT 5.2 (62.2%), and 5.4 nano (46.3%) with GPT-5 (49.6%), but they are far from GPT-5.4’s top performance.
• Computer use: While GPT-5.4 mini achieves an impressive 72.1% in OSWorld‑Verified, which is following GPT-5.4 closely, GPT-5.4 nano is clearly lagging behind (39.0%). It is obvious that it was not created for computer use tasks.

Another thing that struck us immediately was that the order of benchmark scores was the same across (almost) all categories: GPT-5.4 > GPT-5.4 mini > GPT-5.4 nano > GPT-5 mini. Across all published scores, the only exception was that the old mini model beat GPT-5.4 nano in the vision and computer use domains, which are not nano’s target areas.

However, it is unclear how big a difference the new “xhigh” reasoning effort level makes, which wasn’t available for GPT-5 mini.

But of course, performance isn’t everything. OpenAI is trying to make a point about diminishing returns, as shown in the graphs they provided. Among the four models compared, GPT-5.4 is the slowest and most expensive.

The curves illustrate diminishing returns: you can throw more compute/money at a model and get modest accuracy gains, but the jumps get smaller. GPT-5.4's last few percentage points cost a lot more than the first ones. This kind of chart helps engineers decide whether squeezing out that extra 3–4% accuracy is worth the cost in their specific application.

We do feel compelled to evaluate the graph critically, though: The Y-axis starts at 35%, not 0%. This really does exaggerate the visual differences between models. GPT-5.4's lead over GPT-5 mini looks bigger on the chart that starts at 35% than it would when the chart starts at zero. 

Also, OpenAI points out that these latency figures aren't from real production runs; they're modeled estimates. There’s something a little incongruous about this. OpenAI is making a lot of suggestions about infrastructure decisions, and their chart has modeling estimates only. 

We also find it weird to look at modeling estimates without error bars. I would bet the error bars would have overlapped a lot, had OpenAI chosen to include them. 

How to Access GPT-5.4 mini and nano

You should already be able to find GPT‑5.4 mini in the ChatGPT browser UI, Codex, and in the API. In ChatGPT, it is the default “Thinking” model for users in the Free and Go tiers, and the fallback model for all other users who have reached their GPT-5.4 Thinking rate limit.

GPT‑5.4 nano, however, is only available via the API.

GPT-5.4 mini and nano Pricing

GPT-5.4 mini costs $0.75 per 1M input tokens and $4.50 per 1M output tokens. GPT‑5.4 nano, which, again, is only available in the API, costs $0.20 per 1M input tokens and $1.25 per 1M output tokens. For those prices, you get a 400k context window.

That’s obviously much cheaper than OpenAI’s flagship model ($2.50/$15 per 1M input/output). 

GPT-5.4 mini and nano vs. Claude Haiku 4.5

What’s really interesting is that GPT-5.4 nano is priced lower than many lower-performing small models from competitors, namely Claude Haiku 4.5, which is priced at $1 per million input tokens and $5 per million output tokens. So OpenAI undercut the pricing of Claude Haiku in both cases. 

But how do the models compare on the tests? Comparing the two is a little tricky because they have been evaluated on different test variants. The SWE-bench results aren't comparable at all since they use entirely different versions. Claude Haiku 4.5 was tested on SWE-bench Verified and got a score of 73.3%, and GPT-5.4 nano was tested on SWE-bench Pro (Public) and got a score of 52.4%. Pro is the harder and more recent test.

Claude Haiku 4.5 score of 50.7% on OSWorld

On the two cleanest apples-to-apples comparisons, GPT-5.4 nano leads on both.

• With GPQA Diamond, GPT-5.4 nano scores 9.8% higher, and 
• with τ2-bench Telecom, GPT-5.4 nano scores 9.5% higher. 

However, Haiku 4.5 might have an edge on OSWorld computer use, though again, the benchmark variants make the comparison hard. 

• Claude Haiku 4.5 was tested on standard OSWorld and got a score of 50.7%
• GPT-5.4 nano was tested on OSWorld-Verified and got a score of 39.0%. 

OSWorld-Verified is the harder test, but the gap of almost 12% looks fairly significant. We are more cautious about dismissing this gap because, unlike with SWE-bench Verified and SWE-bench Pro, where it’s known that models that do well on the Verified version often do worse on the Pro version, there’s less evidence that the same story applies with OSWorld and OSWorld-Verified.

GPT-5.4 nano score of 39% OSWorld-Verified

What People Are Saying About GPT-5.4 mini and nano

Many online reactions pointed to a familiar pattern in tech: last year's flagship becomes next year's free tier. All this is expected, but the rate of change is startling. 

People are saying frontier AI has the fastest depreciation of any product ever built. People wonder whether the model you're paying a premium for today will still feel worth it in six months. Sometimes, developers might not want to just swap a model for another one if they’ve gone through a process of fine-tuning or if they’ve made some cost and performance calibrations. 

Conclusion

The benchmarks show a clean performance ladder from GPT-5.4 over 5.4 mini to 5.4 nano. But for many tasks, the practical choice depends more on latency and budget than on squeezing out a few extra percentage points. 

For many production apps, GPT-5.4 mini can be a great new default, since its quality is good enough to feel frontier while being cheap and fast enough for high-volume use.

GPT-5.4 nano is more of a specialist for large real-time workloads that are very latency-sensitive. It’s also great for sub-agents to do the easier “mass” work, delegated by higher-performing Thinking models.

In a world where last year’s flagship becomes this year’s “mini”, designing systems that can swap models easily is the superior choice over optimizing for single model releases. I recommend taking our course on Building Scalable Agentic Systems, which addresses this question and teaches you to use agentic frameworks like the Model Context Protocol (MCP).