OpenAI Launches GPT-4.1: A New Era in Language Models

OpenAI has recently unveiled an impressive new lineup of language models, including GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, now accessible through its API. These models represent a significant advancement over their predecessors, GPT-4o and GPT-4.5, showing notable improvements across various technical benchmarks and offering support for an astonishing context of up to 1 million tokens.

Enhanced Performance Across Benchmarks

According to OpenAI, the GPT-4.1 model showcases remarkable enhancements in coding capabilities, instruction adherence, and long-context comprehension. For instance, on the SWE-bench Verified benchmark, which evaluates real-world software engineering tasks, GPT-4.1 achieved an accuracy of 54.6%. This figure reflects a substantial 21-point increase from GPT-4o, which clocked in at 33.2%, and surpasses GPT-4.5 by 26.6 points. Furthermore, in Scale’s MultiChallenge instruction benchmark, GPT-4.1 displayed a 10.5-point improvement over GPT-4o, solidifying its status as a leader in the field.

Long-Context Comprehension

One of the standout features of the GPT-4.1 models is their ability to process extended inputs effectively. Each model in the GPT-4.1 family can manage up to 1 million tokens, a significant leap forward in handling large datasets. Internal evaluations, such as OpenAI-MRCR and Graphwalks, reveal that GPT-4.1 performs exceptionally well in long-context tasks, which require retrieving and reasoning over dispersed information. For example, the model scored 61.7% on Graphwalks, a benchmark designed for multi-hop reasoning, compared to a mere 42% for GPT-4o.

Robust Variants: Mini and Nano

OpenAI has also introduced GPT-4.1 mini and GPT-4.1 nano, designed for specific use cases. GPT-4.1 mini matches or exceeds the performance of GPT-4o in most intelligence evaluations while significantly reducing costs by 83%. On the other hand, GPT-4.1 nano is the most compact and fastest model in the series, tailored for simpler tasks like classification and autocomplete. Despite its size, GPT-4.1 nano achieves impressive scores, such as 80.1% on MMLU and 50.3% on GPQA, showcasing its versatility.

Advancements in Code Editing

Another noteworthy improvement in the GPT-4.1 family is in code editing capabilities. In Aider’s polyglot benchmark, which focuses on the ability to generate diffs rather than full-file rewrites, GPT-4.1 outshines all previous models, including GPT-4.5. The model has drastically reduced the occurrence of unnecessary edits, cutting down from 9% in GPT-4o to just 2% in GPT-4.1, indicating a more refined approach to code generation.

Transition from GPT-4.5 Preview

OpenAI has announced that the GPT-4.5 Preview will be deprecated on July 14, 2025. The company cites the cost and performance improvements in GPT-4.1 as the driving force behind this transition. This decision aligns with community speculation regarding the temporary nature of GPT-4.5. A Reddit user aptly noted that GPT-4.5 served merely as a preview to gather data for developing a more capable and cost-effective model, which ultimately evolved into GPT-4.1.

Pricing Adjustments and Accessibility

Accompanying the launch of the GPT-4.1 family is a revised pricing structure. GPT-4.1 is approximately 26% cheaper than GPT-4o for typical queries, making it more accessible to developers and businesses. Additionally, prompt caching discounts have been significantly increased to 75%, and there are no additional charges for long-context usage beyond standard per-token costs.

Availability and Future Prospects

The GPT-4.1 family is now available via the OpenAI API, although it has yet to be integrated into ChatGPT, where updates to GPT-4o are still underway. This new release not only signifies a leap in AI capabilities but also opens up exciting opportunities for developers and businesses looking to leverage advanced language models for various applications.