GPT-5.4 Mini vs Nano: pricing, benchmarks, and which one to use

What OpenAI actually shipped

On March 17, OpenAI released GPT-5.4 mini and GPT-5.4 nano. Both sit below the full GPT-5.4 ($2.50/$15.00) in the lineup. Mini replaces GPT-5 mini as the go-to mid-tier model. Nano slots in as the cheapest thing OpenAI offers.

Here's what caught our attention: these are not just distilled versions of GPT-5.4. Mini gets computer use support, which is rare for a model at this price point. It scored 72.1% on OSWorld-Verified, nearly matching the full GPT-5.4's 75.0%. That's wild for a model that costs 70% less on input.

Nano is more focused. No computer use, no tool search. It's built for the kind of work where you need a fast answer and don't need the model to browse the web or click buttons. Classification, data extraction, ranking, routing queries to bigger models. API only, not available in ChatGPT.

Full pricing breakdown

Both models support prompt caching and the Batch API. The savings from caching are significant, especially for Nano where cached input tokens drop to $0.02 per million. That's cheaper than most open-source inference providers.

Regional processing endpoints add a 10% surcharge. All prices from OpenAI's API pricing page, retrieved March 23, 2026.

The price hike vs GPT-5 mini and nano

Let's address this directly: GPT-5.4 mini costs 3x more than GPT-5 mini on input and 2.25x more on output. Nano is 4x pricier on input compared to GPT-5 nano. That's not a small jump.

The question is whether the capability jump justifies it. On coding, Mini gains almost 9 points on SWE-Bench Pro (54.4% vs 45.7%). On GPQA Diamond, it jumps from 81.6% to 88.0%. And Mini adds computer use, which GPT-5 mini never had. So the price increase buys real capability. But if your workload ran fine on GPT-5 mini, you might not need it.

Benchmark comparison: Mini vs Nano vs GPT-5.4

We pulled the benchmarks from OpenAI's announcement post. The gaps between Mini and Nano are smaller than you might expect on most tasks, except for computer use where Nano drops off a cliff.

Source: OpenAI announcement

The standout number: Mini hits 72.1% on OSWorld-Verified while Nano drops to 39.0%. That gap is enormous. If you need a model that can operate a browser or desktop, Mini is the only real option between these two. Nano wasn't designed for that.

On everything else, Nano stays surprisingly close. Tool calling (MCP Atlas: 56.1% vs 57.7%), telecom automation (92.5% vs 93.4%), and even coding (52.4% vs 54.4% on SWE-Bench Pro). For tasks that don't require computer use, Nano gives you roughly 95% of Mini's capability at 27% of the price.

How they stack up against the competition

The small model tier got competitive in early 2026. Google's Gemini 2.5 Flash-Lite is aggressively cheap. Claude Haiku 4.5 is more expensive than GPT-5.4 Mini. DeepSeek V3.2 sits in an odd middle ground. Here's the full picture.

If price is all you care about, Gemini 2.5 Flash-Lite and Mistral Small 3.2 are both cheaper than GPT-5.4 Nano on input, and dramatically cheaper on output. Gemini Flash-Lite at $0.40 output vs Nano's $1.25 is a 3x difference.

But price alone doesn't tell you much. GPT-5.4 Mini scores 88.0% on GPQA Diamond. Gemini Flash-Lite isn't competing at that level. If you need reasoning quality at a mid-tier price, Mini is hard to beat. If you need raw volume at the lowest cost, Gemini and Mistral win. You can compare all of these on our pricing page.

What this costs in practice

We ran four common workloads through the TokenCost calculator to see how Mini and Nano compare in monthly spend. Standard API pricing, no caching or batch discounts applied.

With prompt caching enabled, these numbers drop by up to 90% on the input side. If you're sending the same system prompt repeatedly, that changes the math entirely.

When to pick Mini vs Nano

This is simpler than it looks. The decision mostly comes down to whether you need the model to interact with external tools and screens, or just process text.

One thing we noticed: for tool calling that doesn't involve computer use, Nano is surprisingly close to Mini on MCP Atlas (56.1% vs 57.7%). If your agent calls APIs but doesn't need to click around a desktop, Nano saves you 73% on input costs with barely any quality loss.

Speed matters more than you think

OpenAI says both models are "more than 2x faster" than GPT-5 mini. Community measurements suggest Mini runs around 180-190 tokens per second, Nano around 200 tok/s. GPT-5 mini sat at 55-60 tok/s.

For single requests, that difference is a few hundred milliseconds. But in agent pipelines where you chain 5-10 model calls per task, the speed compounds fast. A pipeline that took 30 seconds on GPT-5 mini might finish in 10 seconds on Mini. That's the kind of improvement that changes what you can build in real-time applications. In OpenAI's Codex product, Mini burns only 30% of the GPT-5.4 compute quota, so you can run more parallel tasks before hitting limits.

Sources

• OpenAI: Introducing GPT-5.4 mini and nano (March 17, 2026)
• OpenAI API Pricing (retrieved March 23, 2026)
• OpenAI API Docs: GPT-5.4 mini model card
• OpenAI API Docs: GPT-5.4 nano model card
• The Decoder: GPT-5.4 mini and nano analysis (March 17, 2026)
• Anthropic: Claude API pricing
• Google: Gemini API pricing