📖 Read more: Claude Mythos Leak Exposes Anthropic's Most Dangerous AI Model
🚀 GPT-5.4 Mini: When Speed Beats Size
Two new AI models from OpenAI just shifted the economics of AI development. GPT-5.4 mini and nano landed on March 18, 2026, promising double the speed and dramatically lower costs than previous models. The reason? The age of "subagents" needs tools that can execute specific tasks fast without burning through budgets.
GPT-5.4 mini costs just $0.75 for one million input tokens (about €0.69) and runs over twice as fast as the standard GPT-5.4. Early benchmark testers were stunned — on SWE-bench Pro for software engineering, it hit 54.38%, just 3 percentage points behind its bigger sibling. But the real revolution is in nano.⚡ GPT-5.4 Nano: OpenAI's Cheapest Model Ever
€0.18 Cost per million input tokens
2x Faster than GPT-5.4
Numbers That Surprise
GPT-5.4 mini on the OSWorld-Verified benchmark (measures how well it handles computational tasks) reached 72.13%. Full GPT-5.4 does 75.03%. The gap is so small that in many practical applications, you won't notice it. Nano, as expected, lags more — 39.01% on the same test. But the point isn't to read entire web pages. It's to do its job quickly.🎯 Who They're Built For
OpenAI isn't hiding its intentions. These models target the era of "agentic AI systems" — systems where a large model plans and coordinates while smaller models execute parallel subtasks.Coding Assistants
Fast suggestions, code reviews, real-time debugging without waiting
Code Search
Exploring large codebases, project navigation, targeted edits
Computer Use: The New Frontier
One of the most impressive features is "computer use" — the models' ability to interpret screenshots and execute actions in graphical interfaces. Mini performs excellently here, making it ideal for automation workflows that need to interact with existing software. Imagine an AI assistant that can see an application's interface, understand what it's looking at, and click the right button. This isn't science fiction — it's available now.💰 Economics and Availability
GPT-5.4 Mini Pricing:
• Input: €0.69 per million tokens
• Output: €4.14 per million tokens
• Cached input: €0.069 per million tokens
GPT-5.4 Nano Pricing:
• Input: €0.18 per million tokens
• Output: €1.15 per million tokens
Mini is available everywhere: ChatGPT (including free users), API, Codex. Nano, however, is API-only — a strategic choice to keep it focused on enterprise applications. At Microsoft Foundry, which is OpenAI's key partner, the models launched simultaneously. This means enterprise customers can test them immediately in production environments.• Input: €0.69 per million tokens
• Output: €4.14 per million tokens
• Cached input: €0.069 per million tokens
GPT-5.4 Nano Pricing:
• Input: €0.18 per million tokens
• Output: €1.15 per million tokens
The Competition Comparison
OpenAI isn't alone in this move. Anthropic's Claude 4.5 Haiku targets the same lightweight agent tasks. Google's Gemini 3 Flash too. But OpenAI's pricing and integration with its ecosystem (Codex, ChatGPT) gives it an edge. What's interesting is that this "AI race to the bottom" in pricing terms doesn't mean worse quality. Instead, we're seeing better specialization.🔬 Benchmarks and Real Performance
Official benchmarks are impressive, but what do early users say? Abhisek Modi from Notion AI Engineering puts it clearly: "GPT-5.4 mini handles focused, well-defined tasks with impressive accuracy. For page editing specifically, it matches and often exceeds GPT-5.2 on complex formatting with a fraction of the compute."This is critical. Reliability in tool calling was the last barrier to mass adoption of AI agents. If small models can now call APIs and functions with the same reliability, then the economic equation changes radically."Until recently, only the most expensive models could reliably handle agentic tool calling. Today, smaller models like GPT-5.4 mini and nano do it easily."
Abhisek Modi, Notion AI Engineering
Where They Fall Short
Not everything is rosy. Nano on OSWorld-Verified (39.01%) lags behind the old GPT-5 mini (42%). Both models can't handle long-term, complex analyses like the main GPT-5.4. But that's the point — they weren't designed for that. They were designed to do specific things well, fast and cheap.🌐 What It Means for Global Developers
The release of these models is a breakthrough for the global startup ecosystem. Small companies that couldn't afford GPT-5.4 costs for high-volume applications now have options. Imagine an application that categorizes thousands of emails daily for a company. With nano at €0.18 per million tokens, the cost becomes negligible. At the same time, quality remains high.Cost example: 10,000 emails daily with an average of 200 tokens each = 2 million tokens/month = €3.60/month for categorization with nano
These numbers open the conversation for AI-powered features in applications that didn't have them before because the cost was prohibitive.🎯 Frequently Asked Questions
How different are they from the old mini/nano models?
The last mini/nano were based on GPT-5 (2025). The new ones are based on GPT-5.4 and bring significant improvements especially in code, computer use and multimodal understanding. Speed is also double.
Can I use mini for free in ChatGPT?
Yes, GPT-5.4 mini is available to free ChatGPT users from March 18, 2026. Nano is API-only.
Which model should I choose for my application?
If you want fast reasoning with good quality (coding assistants, real-time agents), take mini. If you need high throughput for simple tasks (classification, extraction), take nano. For complex analysis, stick with main GPT-5.4.
What becomes clear is that 2026 is no longer the era of "one model for everything." It's the era of the right tool for the right job. And with these new models, OpenAI makes the multi-model approach economically feasible for everyone. It is uncertain what developers will build with these tools.Sources: