OpenAI released GPT-5.4 mini and nano on March 17, 2026, calling them its “most capable small models yet.” These aren’t stripped-down versions.

They’re purpose-built for speed and cost. GPT-5.4 mini significantly improves over GPT-5 mini across coding, reasoning, multimodal understanding, and tool use, while running more than 2x faster. GPT-5.4 nano is the smallest, cheapest version designed for tasks where speed and cost matter most. If your team runs high-volume

AI workloads, you’re about to cut costs dramatically while maintaining quality. In this post, we’ll explain what happened, how these models perform, and exactly when to use each one.

The Release: Two New Models Built For Scale

OpenAI released two new models yesterday designed specifically for developers building agentic AI systems. The bigger problem they’re solving: most teams route every task through expensive flagship models when cheaper, faster alternatives handle 80% of the work just fine.

GPT-5.4 mini is available in ChatGPT, Codex, and the API. For ChatGPT users, Free and Go users can access it via the “Thinking” feature in the plus menu. Paid subscribers who hit rate limits automatically fall back to mini. GPT-5.4 nano is API-only, costing $0.20 per million input tokens and $1.25 per million output tokens.

The pricing tells you everything. Nano is roughly four times cheaper than mini on inputs, making it financially realistic for startups to run huge volumes of queries per day. This is the kind of price point that changes what you can build.

Why This Matters: The Agentic Architecture Shift

Here’s where this gets interesting. These models are built for workloads where latency directly shapes product experience: coding assistants that need to feel responsive, subagents that quickly complete supporting tasks, and computer-using systems.

Think about how AI workflows actually run at scale. You don’t need the flagship model to read a document, extract specific data, or run a simple classification. You need it to plan what to extract and coordinate the work. The heavy lifting gets delegated to smaller models running in parallel.

GPT-5.4 mini approaches GPT-5.4 on coding benchmarks at a fraction of the cost and is designed to be delegated to, not used alone. This is the first time OpenAI has launched new mini and nano models for production agentic systems.

How This Changes Your Budget

The math changed overnight. GPT-5.4 mini is more than two times faster than GPT-5 mini, which matters when you’re paying $0.75 per million input tokens instead of premium rates.

Scenario 1: You run a coding assistant. Mini uses only 30% of the GPT-5.4 quota in Codex while handling routine coding tasks. Your cost drops 70% for tasks that don’t need the flagship model. You still get 72.1% accuracy on OSWorld-Verified computer use tasks compared to the flagship’s 75%.

Scenario 2: You’re building a customer support system. If you’re running a customer service chatbot answering the same 200 questions daily, you don’t need the model that scores best on PhD-level chemistry exams. Nano handles this at one-quarter mini’s cost.

Scenario 3: You process data at scale. Classification, data extraction, ranking, and coding subagents handling simpler supporting tasks are nano’s sweet spot. Running 100,000 extractions per day now costs $52 instead of thousands of dollars.

The trap teams fall into: assuming you need the expensive model everywhere. OpenAI is explicitly pushing a tiered architecture where GPT-5.4 handles planning and complex judgment while smaller models execute narrower tasks in parallel. This is how you scale without burning budget.

What You Can/Could/Should Do Now

1. Audit your current GPT-5.4 usage. How much of your spend goes to tasks that don’t actually require GPT-5.4’s reasoning ability? Data extraction. Formatting. Simple classifications. Those tasks move to nano immediately.

2. Test mini on your coding workloads. On SWE-Bench Pro, mini scored 54.4% compared to the flagship’s 57.7%—a narrow gap that matters when costs drop significantly. If the quality difference doesn’t impact your users, the cost savings transform your margins.

3. Build for delegation, not single models. Instead of routing every task through an expensive flagship model, you can build systems where the big model plans and coordinates while smaller models handle actual grunt work in parallel. This architecture unlocks economies of scale.

4. Understand when nano isn’t enough. Nano trades capability for cost efficiency. On OSWorld-Verified computer-use tasks, it scored 39% versus mini’s 72%. If you need computer vision or complex reasoning, nano will disappoint. Use mini instead.

Practical starting point: Move 30% of your current workloads to mini. Move another 20% to nano. Measure quality. If performance holds, that’s a 40%+ cost reduction on those workloads. Scale from there.

Closing

This isn’t a minor update. OpenAI describes GPT-5.4 mini as approaching the performance of the larger GPT-5.4 model on several evaluations, including SWE-Bench Pro and OSWorld-Verified. You get near-flagship performance at one-fifth the cost, plus 2x speed.

Teams that restructure their AI systems to use tiered architectures this month will have massive cost advantages by Q3. Teams that don’t will continue overspending on tasks that don’t need premium models. The gap only widens from here.

Check the official OpenAI announcement for the full technical details and benchmark comparisons. Your AI budget conversation just changed. The question isn’t whether you’ll migrate to mini and nano. It’s how quickly your team can make the transition.