OpenAI Launches Cheaper AI Models Nearing Flagship Power

GPT-5.4 Mini Narrows Performance Gap to 3.3% at 70% Lower Cost

On March 17, OpenAI launched two new smaller artificial intelligence models, GPT-5.4 mini and nano, fundamentally altering the cost-performance equation for developers. The GPT-5.4 mini model delivers performance that closely rivals its flagship counterpart, GPT-5.4, but at a significantly lower price point. The mini's API pricing is set at $0.75 per million input tokens and $4.50 per million output tokens, representing a 70% cost reduction compared to the flagship's respective rates of $2.50 and $15.00.

This cost efficiency does not come with a major trade-off in capability. On the SWE-bench Pro benchmark, which tests the ability to solve real-world GitHub issues, GPT-5.4 mini achieved a score of 54.4%, just 3.3 percentage points shy of the flagship GPT-5.4's 57.7%. In computer operation tasks measured by the OSWorld-Verified benchmark, the mini model scored 72.1%, nearly matching the flagship's 75.0% and slightly exceeding the human baseline of 72.4%.

New Models Enable 'Sub-Agent' Architecture for AI Workflows

The release signals a strategic push towards multi-model systems where a powerful AI acts as a coordinator for smaller, faster models. This 'sub-agent' architecture allows developers to delegate high-volume, less complex tasks to cheaper models, reserving the expensive flagship AI for high-level planning and reasoning. For example, within OpenAI's own Codex programming assistant, GPT-5.4 can orchestrate tasks while GPT-5.4 mini sub-agents handle parallel jobs like code searches and file reviews at about one-third the cost.

This approach is critical for latency-sensitive applications where response speed directly impacts user experience. The new models' ability to handle complex tool-calling reliably at a low cost is a significant step forward for the industry.

Until recently, only the most expensive models could reliably navigate agentic tool calling. Today, smaller models like GPT-5.4 mini and nano can easily handle it, which will let our users build Custom Agents on Notion pick exactly the amount of intelligence they need.

— Abhisek Modi, AI Engineering Lead at Notion.

GPT-5.4 Nano Targets High-Volume Tasks at $0.20 per Million Tokens

Positioned as the most economical option, GPT-5.4 nano is available exclusively through OpenAI's API and is priced at just $0.20 per million input tokens and $1.25 per million output tokens. OpenAI designed this model specifically for high-frequency, low-complexity tasks such as data classification, text extraction, and simple coding support.

While nano's performance on visual tasks like OSWorld-Verified is lower than its predecessors, it demonstrates a significant improvement over previous small models in reasoning and tool-use benchmarks. For instance, its score of 82.8% on the GPQA Diamond general intelligence test surpasses the older GPT-5 mini's 81.6%. The introduction of nano provides developers with a highly specialized tool for building cost-effective AI systems where speed and budget are the primary constraints.