OpenAI has just launched three new models within its API: GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. These models bring significant improvements in coding, instruction following, and offer extended context management, up to 1 million tokens.

Optimized performance

The flagship model, GPT-4.1, significantly enhances performance compared to GPT-4o, particularly in the following areas:

  • Coding: GPT-4.1 achieves 54.6% on the SWE-bench Verified benchmark, a notable increase from GPT-4o (33.2%) and GPT-4.5 (38%). This result reflects an improvement in its ability to solve complex software development problems.

  • Instruction following: On the Scale MultiChallenge evaluation, GPT-4.1 scores 38.3%, improving by 10.5 absolute points compared to GPT-4o.

  • Long context: GPT-4.1 sets a new record with 72% on the Video-MME benchmark, dedicated to understanding multimodal content in an extended context.

A complete range to meet varied needs

In addition to GPT-4.1, the mini and nano variants offer effective compromises between performance and cost:

  • GPT-4.1 mini: This more compact model surpasses GPT-4o in several evaluations while significantly reducing latency (nearly by half) and cost (reduction of 83%).

  • GPT-4.1 nano: The fastest and most economical model, ideal for tasks such as classification or autocompletion, offering despite its reduced size, a context up to 1 million tokens.

Enhanced capabilities for intelligent agents

Thanks to its improvements in instruction-following reliability and extended context understanding, GPT-4.1 strengthens applications based on autonomous agents. Developers can now build more reliable and efficient systems for document management, software development, or automated customer request processing.

Planned end of GPT-4.5 Preview

OpenAI has announced the upcoming depreciation of the GPT-4.5 Preview model in favor of GPT-4.1, offering superior performance at lower cost. GPT-4.5 Preview will be deactivated starting July 14, 2025, to allow developers to make a smooth transition.

Exclusive availability via API

It should be noted that GPT-4.1 will be exclusively available through the OpenAI API. However, ChatGPT users will gradually benefit from the improvements of GPT-4.1 integrated into the GPT-4o version.

Optimized pricing

With pricing revised downwards, GPT-4.1 is now available at a cost 26% lower than GPT-4o for common requests. The highly competitive pricing of GPT-4.1 nano makes it the most affordable offer ever proposed by OpenAI.

 

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

Translated from OpenAI annonce le lancement de GPT-4.1 : accent sur le codage, le suivi des instructions et la gestion étendue du contexte

To better understand

What is the SWE-bench Verified benchmark used to evaluate OpenAI's models?

The SWE-bench Verified is a suite of tests designed to assess AI models' abilities to solve complex software development problems. It measures the models' coding skills and their efficiency in following precise development instructions.

What is the current regulation regarding AI models like GPT-4.1?

AI models like GPT-4.1 must comply with regulations that include personal data protection, algorithm transparency, and accountability for bias. The European Union is working on the AI Act, which could impose strict compliance standards for commercially used models.