GPT-5.4 Mini and Nano: 2x Faster OpenAI Models Launched

Six months after GPT-5.4 dropped, OpenAI just fired back with two models built to change AI economics. GPT-5.4 Mini and Nano aren't just cheaper versions of the flagship. They're purpose-built for the "agentic economy" — AI models delegating tasks to other AI models.

On March 17, 2026, OpenAI announced models that run fast, cost reasonable money, and don't sacrifice much capability. GPT-5.4 Mini runs over 2x faster than its predecessor and scores 54.4% on SWE-Bench Pro — just 3.3% below the flagship GPT-5.4. Nano costs only €0.20 per million tokens and was built for the grunt work nobody wants: classification, extraction, ranking.

🚀 GPT-5.4 Mini: The Model That Didn't Compromise

Let's start with the numbers that actually matter. GPT-5.4 Mini isn't GPT-5 Mini with better branding. It's a completely new architectural approach targeting that sweet spot where speed meets capability. In coding, Mini scores 54.4% on SWE-Bench Pro. For reference, GPT-5 Mini hit 45.7% and the flagship GPT-5.4 reaches 57.7%. Mini operates at 94% of the flagship's coding capability, but costs and runs fast enough for production use.

🎯 Benchmark Performance That Counts

SWE-Bench Pro (Coding): 54.4% vs 45.7% for GPT-5 Mini
OSWorld-Verified (Computer Use): 72.1% vs 42.0% for GPT-5 Mini
GPQA Diamond (Reasoning): 88% vs 81.6% for GPT-5 Mini
Toolathlon (Tool Use): 42.9% vs 26.9% for GPT-5 Mini

The most impressive jump is in OSWorld-Verified benchmarks. We're talking about the model's ability to "see" screenshots and execute computer-use tasks. Mini scores 72.1% versus GPT-5 Mini's 42.0% — a 30-point leap that changes the game for anyone building automation tools.

Where Mini Drops Points

Not everything is perfect. In long-context needle retrieval tasks, Mini scores 47.7% versus the flagship GPT-5.4's 86%. If your work involves finding specific facts in very long documents, Mini will disappoint you. It's one of the few cases where the gap with the flagship is dramatic — and worth knowing before you migrate workloads.

💰 GPT-5.4 Nano: The €0.20 Model

Nano is something completely different. It's not "cheap Mini" — it's a model designed from scratch for subagent workflows and bulk processing tasks. At €0.20 per million input tokens (with cached input at €0.02), Nano changes the economics of agentic pipelines. Simon Willison ran the math: describing 76,000 photos costs about €49. For media companies or e-commerce platforms, that's workable production cost.

52.4% SWE-Bench Pro Score

€0.18 Per Million Tokens

39.0% OSWorld Computer Use

Nano scores 52.4% on SWE-Bench Pro, beating GPT-5 Mini's 45.7%. The smallest in the new family outperforms the mid-tier of the previous generation in coding. This isn't incremental improvement — it's an architectural leap.

What Nano Can't Do

In OSWorld-Verified, Nano scores just 39%, slightly below even GPT-5 Mini. For computer-use tasks, Nano is off-limits. That's the basic difference that will determine when you use Mini versus Nano.

🔧 The Architecture of the Multi-Model Era

The real innovation isn't the models per se — it's the workflow pattern OpenAI is proposing. Orchestrator models (GPT-5.4) handle planning and complex decisions. Execution models (GPT-5.4 Mini) run mid-complexity subtasks with speed. Worker models (GPT-5.4 Nano) do the "dirty work" — classification, extraction, ranking.

"Larger models handle planning and judgment, while GPT-5.4 Mini executes narrower subtasks"
OpenAI Documentation

This architecture solves a fundamental economics problem the AI industry had. Expensive models were too costly for subagent frequency. Cheap models were too weak to be reliable. Mini and Nano together close that gap.

🎮 What Changes for Developers

The immediate impact will hit coding assistant developers first. Mini scores 42.9% on the Toolathlon benchmark (tool-calling accuracy) versus GPT-5 Mini's 26.9% — a 60% relative improvement. In agentic pipelines, tool-calling errors compound exponentially. One wrong function call can destroy an entire workflow. In the Codex environment, Mini uses only 30% of GPT-5.4's quota. That means roughly three times more tasks with the same budget allocation — a breakthrough for teams running automated code reviews or pull request generation.

Free Tier Users: The Instant Upgrade

Starting March 17, free ChatGPT users get access to GPT-5.4 Mini through "Thinking" mode. This means they're getting a model that outperforms GPT-5 Mini on all major benchmarks — without changing their subscription cost.

For Coding

54.4% SWE-Bench Pro score — real upgrade for coding assistants and automated debugging

For Computer Use

72.1% OSWorld score — very close to flagship for UI automation and screenshot interpretation

For Reasoning

88% GPQA Diamond — doctoral-level reasoning at a price that makes sense for production

⚡ Speed vs Capability: The New Paradigm

OpenAI is focusing on something the industry underestimated: in real-world applications, fast models often outperform larger ones because they respond instantly and remain capable enough for complex tasks. This is especially important for real-time environments — coding assistants that need instant feedback, subagents handling background tasks, systems analyzing screenshots. Latency directly shapes user experience, and Mini gets this.

🏁 Context Window: The Silent Upgrade

Both models support a 400,000-token context window. This might not sound like a big deal, but it changes the economics of agent workflows. Agents often accumulate massive context windows through conversation history, retrieved documents, tool call results. Before, this would break at 128K tokens and require expensive context trimming operations. At 400K, this overhead disappears — agents can maintain richer state without needing constant management.

🎯 Frequently Asked Questions

Which model should I choose for coding tasks?

For planning and architecture decisions, stick with GPT-5.4. For code review, debugging, and automated testing, Mini is the perfect fit. For classifying bugs or extracting metadata from codebases, Nano is what you need.

Can I use Nano for computer-use automation?

No. The 39% OSWorld score makes it unsuitable for UI interpretation. If you need screenshot analysis or browser automation, Mini (72.1%) is the right choice.

What does cached input pricing mean?

If you reuse the same input text across multiple API calls, cached input costs only €0.02/M tokens for Nano. For bulk processing workflows, this can reduce costs by 90%. These models determine which AI applications can scale economically. GPT-5.4 Mini and Nano represent OpenAI's bet that AI's future is multi-model, specialized, and affordable.

GPT-5.4 OpenAI AI models machine learning artificial intelligence coding assistants AI agents model optimization

Sources:

OpenAI Launches GPT-5.4 Mini and Nano: Purpose-Built Models for the Agentic Economy