Apr 15, 2025 11:15:00

OpenAI releases 'GPT 4.1' API, next-generation AI model with 1 million token support and practical performance

OpenAI announced the APIs for ' GPT-4.1 ', ' GPT-4.1 mini ', and ' GPT-4.1 nano ' on April 14, 2025. The GPT-4.1 series supports input of up to 1 million tokens, and its coding and instruction tracking capabilities have been significantly improved compared to the GPT-4o series, making it a low-latency, high-cost-performance model.

Introducing GPT-4.1 in the API | OpenAI

https://openai.com/index/gpt-4-1/

GPT-4.1 (and -mini and -nano) are now available in the API!

These models are great at coding, instruction following, and long context (1 million tokens).

benchmarks are strong, but we focused on real-world utility, and developers seem very happy.

GPT-4.1 family is API-only.

— Sam Altman (@sama) April 14, 2025

OpenAI's announcement of GPT-4.1 can be seen in the following movie.

GPT 4.1 in the API - YouTube


A feature of GPT-4.1 is that it supports a very long context window of up to 1 million tokens. This makes it possible to process hundreds of pages of documents and large code bases at once and make appropriate inferences while retaining the content. OpenAI says that this ability to process long texts is particularly important in many practical fields, such as law, finance, and customer support.

In terms of coding, GPT-4.1 achieved a score of 54.6% in

SWE-bench Verified , which measures coding ability by solving issues collected from GitHub, far exceeding GPT-4o's 33.2%. It also recorded high accuracy in generating code differences across multiple languages, reducing development costs and latency by outputting only the changes rather than the entire file.

In

Scale's MultiChallenge , a benchmark that measures the ability to follow instructions, GPT-4.1 scored 38.3%, which is 10.5 percentage points better than GPT-4o. Furthermore, in Video-MME , a benchmark for multimodal long-term context understanding, GPT-4.1 scored 72.0% in the 'long-term, no subtitles' category, an improvement of 6.7 percentage points over GPT-4o. OpenAI reports that GPT-4.1 is now able to respond more accurately than before in a variety of instruction categories, such as specifying the format, prohibitions, output order, and including specific information, and that there was a greater improvement than GPT-4o in particularly difficult prompts.

Evolution has also been seen in the understanding of visual information, scoring highly in visual question-answering tasks that include diagrams and mathematical formulas. In particular, the GPT-4.1 mini model is lighter than GPT-4o, yet is increasingly outperforming it in image processing capabilities, maintaining high performance while reducing costs and latency.

In addition, response speed and cost performance have also been improved. GPT-4.1 is up to 26% cheaper than previous models, and the reduction rate due to prompt cache has been expanded to 75%. Lightweight 'mini' and 'nano' models are also available at the same time, and can be used for applications that require low latency and cost optimization.

These characteristics make GPT-4.1 a practical AI model that enables the automation of complex tasks and the construction of intelligent agents beyond previous limits, and is optimized for use via APIs. It is expected that future releases will continue in this direction and meet even more diverse development needs.

At the time of writing, the GPT-4.1 series is available on the OpenAI API platform and is not available for use with ChatGPT. However, ChatGPT's GPT-4o is gradually incorporating improvements introduced in GPT-4.1, such as instruction following and code generation, and further integration is planned for the future.

The API prices are as follows:

In addition, with the advent of GPT-4.1, the previous 'GPT-4.5 Preview' model has been decided to be discontinued on the API on July 14, 2025, and developers are being given a transition period. GPT-4.5 was initially provided as a research preview that used a lot of computational resources, and since it has been confirmed that GPT-4.1 is superior in many aspects, it will be gradually replaced.