OpenAI has announced a new family of models, GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, which the company claims outperform their predecessors, GPT-4o and GPT-4o mini, "across the board." The new models are available only via the API and offer significantly larger context windows, improved long-context comprehension, and increased output token limits.

In conjunction with the launch of the GPT-4.1 family, OpenAI is deprecating GPT-4.5 Preview in the API, which will be turned off completely on July 14, 2025. The company cites the new models' similar or better performance for many functions at lower cost and latency as the reason for the deprecation.

The GPT-4.1 models boast a significantly larger context window of one million tokens, compared to GPT-4o's 128,000 tokens. This improvement enables the models to better comprehend long-context tasks and perform more complex functions. Additionally, the output token limits have been increased from 16,385 in GPT-4o to 32,767 in GPT-4.1.

OpenAI worked closely with the developer community to optimize the models to meet their priorities. For example, the company improved the coding score on SWE-bench by 21.4% over that of GPT-4o. The GPT-4.1 mini and GPT-4.1 nano models are specifically touted for their performance, with the mini model beating GPT-4o in many benchmarks and reducing latency by nearly half and cost by 83%.

The GPT-4.1 nano model is ideal for tasks that demand low latency, delivering exceptional performance at a small size with its one million token context window. It scores 80.1% on MMLU, 50.3% on GPQA, and 9.8% on Aider polyglot coding, even higher than GPT-4o mini.

OpenAI claims that the improvements in the GPT-4.1 family, combined with primitives such as the Responses API, will allow developers to build more useful and reliable agents that can perform complex tasks such as extracting insights from large documents and resolving customer requests "with minimal hand-holding."

The new models are also more cost-effective, with a 26% lower cost than GPT-4o for median queries. The prompt caching discount is increasing from 50% to 75%, and long context requests are billed at the standard per-token price. Additionally, the models can be used in OpenAI's Batch API at an additional 50% discount.

However, some analysts have raised questions about the claims. Justin St-Maurice, technical counselor at Info-Tech Research Group, notes that the announcement brings up questions about efficiency, pricing, and scale. He questions the baseline or model being compared to achieve the 83% cost reduction and notes that the models are premium offerings.

St-Maurice believes that OpenAI's focus on long-context performance and more efficient variants like mini or nano aligns with current conversations around Model Context Protocol servers and agentic systems. However, he emphasizes the need for more transparency with practical benchmarks and pricing baselines to strengthen OpenAI's position for efficient, scalable intelligence.

Overall, the GPT-4.1 family represents a significant leap forward in AI performance and cost-effectiveness, with potential applications in complex tasks such as coding, document analysis, and customer service. As the AI landscape continues to evolve, it remains to be seen how OpenAI's new models will be adopted and utilized by developers and enterprises.