OpenAI has launched three new artificial intelligence models: GPT‑4.1, GPT‑4.1 mini, and GPT‑4.1 nano. These enhanced versions surpass their predecessors in code generation, instruction tracking, and long-context understanding, while also offering greater efficiency and reduced costs. With an expanded context window of up to 1 million tokens and a knowledge base updated through June 2024, these models are available exclusively through the company’s API.
The main model, GPT‑4.1, stands out for its performance in coding tasks, exceeding GPT‑4o by 21% and GPT‑4.5 by 27% in standard evaluations. The mini and nano versions offer more economical and faster solutions, maintaining a high level of performance, making them ideal for a wide range of applications, from autonomous agents to analysis of large volumes of data.
Below, we analyze the performance of the three models in different areas, with real examples.
GPT‑4.1 has been specifically optimized to overcome real challenges in software engineering. In the SWE-bench Verified evaluation (a metric that assesses the ability to resolve tasks from incident descriptions in code repositories, generating functional and validated patches), this model achieved 54.6% effectiveness, compared to GPT‑4o’s 33.2%, representing an absolute improvement of 21.4%.
Among the highlighted improvements, GPT‑4.1 demonstrates greater consistency in using tools, better tracking of diff formats, and a notable reduction in unnecessary edits. These skills make it more reliable for tasks like frontend coding, agent-based development, and refactoring of large files (a common practice in software development to reorganize and optimize code within extensive files without changing its external functionality).
GPT‑4.1 mini, despite its reduced size, equals or surpasses GPT‑4o in intelligence evaluations, reducing latency almost by half and cost by 83%.
Finally, GPT‑4.1 nano, the fastest and most economical model, achieved a score of 9.8% in the Aider polyglot coding test (a benchmark designed to evaluate the capability of LLMs in editing and generating code in multiple programming languages), surpassing GPT‑4o mini, which is impressive considering its size and speed. Additionally, it achieved a score of 80.1% in MMLU (measuring language comprehension and multitask knowledge) and 50.3% in GPQA (measuring deep reasoning capability in physics), making it an ideal option for tasks like classification and autocompletion.
These improvements enable developers to create more efficient applications, from coding assistants to automated code review systems, facilitating the integration of artificial intelligence into software development processes.
For example, during a demonstration, GPT‑4.1 was able to develop an educational flashcard application for language learning, following detailed instructions and adapting to the user’s specific needs.
GPT‑4.1 significantly enhances its ability to follow instructions, greatly outperforming earlier versions like GPT‑4o, especially in complex tasks. This improvement has been measured through an internal evaluation developed by OpenAI, focusing on 6 areas:
Each category is assessed at difficulty levels of easy, medium, and hard. GPT‑4.1 excels especially at the hard levels, with 49% accuracy compared to GPT‑4o’s 29%.
In the benchmark Scale’s MultiChallenge (evaluation developed by Scale AI to measure the ability to maintain realistic, complex conversations with multiple interaction turns), GPT‑4.1 scored 38.3%, a 10.5% jump compared to GPT‑4o. This improvement translates into more natural interactions, reducing the need to rephrase instructions.
This is particularly useful for building autonomous agents that interact with users or systems, such as virtual assistants, issue resolution systems, or creative writing agents.
For example, companies such as Blue J and Qodo (specialized in code generation) have implemented GPT‑4.1 to generate legal documents and academic summaries with precision, while the editing tool Windsurf and the agency Thomson Reuters use it in development environments where every instruction must be followed to the letter.
All models from the GPT‑4.1 family support a context window of 1 million tokens, enabling the processing of extremely long text files. In the Vídeo-MME evaluation, GPT‑4.1 achieved 72% in the “long, no subtitles” category, a 6.7% improvement over GPT‑4o.
This allows for new use cases such as the complete analysis of technical documentation, reading financial reports hundreds of pages long, or the understanding of video scripts or subtitles with a single query, identifying key trends and providing understandable analysis for strategic decision-making.
Furthermore, thanks to this capability, it is possible to build agents that operate autonomously for hours, maintaining coherence and accuracy without needing to reset the context.
Although the GPT‑4.1 models are not multimodal like GPT‑4o, they have been designed to integrate effectively with image generation tools, enhancing coherence between text and image. This enables developers to create applications that combine detailed textual descriptions with accurate visual representations, facilitating the creation of attractive and personalized multimedia content.
For example, an application can use GPT‑4.1 to generate detailed product descriptions and then, through an image generation tool, create visual representations that match these descriptions, improving user experience and marketing effectiveness.
Photo: OpenAI
Your email address will not be published. Required fields are marked *
Δ