OpenAI Releases GPT-4.1 Family of AI Models in API, Said to Outperform GPT-4o

OpenAI released three new artificial intelligence (AI) models on Monday. Staying true to its unusual naming convention, the company called this the GPT-4.1 family of AI models. There are three variants — GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. The San Francisco-based AI firm said that these models outperform the GPT-4o and GPT-4o mini across several metrics, with a significant improvement in coding and instruction following. Notably, these new large language models (LLMs) are only available to developers via an application programming interface (API), and will not be accessible via ChatGPT.

OpenAI Releases Coding-Focused GPT-4.1 AI Models

In a post, OpenAI announced the release of the GPT-4.1 series of AI models. These API-exclusive models get several upgrades over the GPT-4o models, including higher performance, increased context window, and improved pricing. The knowledge cutoff of these models has also been upgraded to June 2024.

OpenAI said that these models were designed for coding-related tasks and actions that require higher instruction adherence. Highlighting the coding proficiency, OpenAI claimed GPT-4.1 scored 54.6 percent on the SWE-bench Verified. This score is higher than both GPT-4o and the recently released GPT-4.5, making 4.1 the company's leading model for coding.

GPT-4.1 also scored 38.3 percent on the MultiChallenge benchmark for instruction following, and 72 percent on the Video-MME benchmark for multimodal long context understanding, the company claimed. Notably, all the AI models in the series have a context window of up to one million tokens.

Overview of OpenAI's GPT 4.1 AI models
Photo Credit: OpenAI

While GPT-4.1 remains the premium model of this series, the GPT-4.1 mini is also claimed to outperform GPT-4o in several benchmarks, including intelligence evaluations. OpenAI said the mini model reduces latency by half and costs 83 percent less compared to 4o.

The AI firm is also pitching the GPT-4.1 nano as the best model for low latency tasks. With an 80.1 percent score on MMLU, 50.3 percent score on GPQA, and 9.8 percent on the Aider polyglot coding, OpenAI says the AI model fares better than the GPT-4o mini model. OpenAI also highlighted that the improved instruction following and the long context window also make these models an option in powering AI agents.

Coming to pricing, the GPT-4.1 costs $2 (roughly Rs. 171) per million input tokens and $8 (roughly Rs. 685) per million output tokens. The mini variant is priced at $0.40 (roughly Rs. 34) per million input tokens and $1.60 (roughly Rs. 137) per million output tokens. And, the cheapest nano model will cost developers $0.10 (roughly Rs. 8.5) per million input tokens and $0.40 (roughly Rs. 34) per million output tokens.

Notably, the GPT-4.1 series of AI models will only be available via API to developers. However, the AI firm stated that the improvements of these models will be transitioned into the latest version of GPT-4o, which is available in ChatGPT.