With GPT-4.1 and its mini and nano versions, OpenAI offers, via its API, three new models with improved skills, particularly in terms of coding.
On April 14, 2025, OpenAI unveiled a new family of artificial intelligence models called GPT-4.1, consisting of three distinct models. Accessible only via API, they are expected to offer significant improvements in coding, instruction following, and understanding extended contexts.
GPT-4.1: three new models accessible via the API
According to OpenAI, these three new models, dubbed GPT-4.1, GPT-4.1 mini and GPT-4.1 nano, “excel at coding tasks and following instructions .” Their million tokens allow them to process around 750,000 words in a single query while having been “optimized (…) to improve certain priority areas for developers: front-end coding, respect for formats, structure, order of responses, consistent use of tools and much more .” Improvements “that allow developers to create significantly more efficient agents for real-world software engineering tasks,” an OpenAI spokesperson told US media.
Here are the specific features of these three GPT-4.1 models:
- GPT-4.1: The most comprehensive model, with the best performance for coding, understanding complex contexts, and advanced tasks. It can handle up to 1 million tokens and is suitable for demanding use cases.
- GPT-4.1 mini: A lightweight version that’s 83% cheaper than GPT-4o, while maintaining high efficiency. Ideal for high-performance, low-cost AI tools, especially in development.
- GPT-4.1 nano: Ultra-fast and cost-effective, designed for simple, high-frequency tasks such as classification or autocompletion. Suitable for applications requiring low latency at large scale.
These models outperform GPT-4o and GPT-4o mini in every way, with major gains in coding and instruction tracking, OpenAI boasts.
Significantly improved performance, particularly for code
According to benchmarks shared by OpenAI, GPT-4.1 models show significant progress over their predecessors, particularly on software development tasks. The main model, GPT-4.1, achieved a score of 54.6% on the SWE-bench Verified benchmark, which evaluates an agent’s ability to resolve GitHub issues in open-source projects. By comparison, GPT-4o achieved a score of 33.2%. This is a significant improvement, but still far from a standalone level for production software development.
This improvement is reportedly due to targeted optimization work focused on developer expectations, particularly in terms of structural rigor in responses, consistency in the use of tools (such as APIs or specific languages), and technical project management. OpenAI also states that the model is more reliable in following precise instructions, which would make it a better assistant for tasks such as code refactoring, unit test generation, or automated documentation.
Finally, the ability of these models to process up to a million tokens in a single query will allow users to submit large corpora, such as complete code bases, technical documents or extensive conversation histories.
API-designed models with competitive pricing
GPT-4.1 models will not be available in the ChatGPT interface, even for paid subscribers, in part because “the latest version of 4o in ChatGPT incorporates many of the same improvements .” They are therefore exclusively accessible through the OpenAI API, which reserves them for uses integrated into tools or services developed by third parties, and therefore mainly for developers and companies able to integrate them into their own products or workflows.
In terms of pricing, OpenAI has divided its models into three price levels to accommodate different use cases. The main model, GPT-4.1, is priced at $2 per million tokens for input and $8 for output. GPT-4.1 mini, which is lighter but powerful, costs $0.40 (input) and $1.60 (output). As for GPT-4.1 nano, designed for simple tasks at scale, it starts at $0.10 (input) and $0.40 (output). This pricing, which is intended to be competitive with unprecedented prices for models presented as being so powerful, was designed to encourage adoption: GPT-4.1 mini is in fact 83% cheaper than GPT-4o, and GPT-4.1 is 26% less expensive “for median queries.“
