文档 图片 29/10000 实时翻译 划译 OpenAI's Big Shot! GPT-4.1 is launched, with full programming power

On April 15th, the US artificial intelligence company OpenAI announced on Monday the launch of a series of new artificial intelligence models, which are optimized for programming tasks and aim to counter the increasingly fierce competition from companies such as Google and Anthropic. Developers can now use these models through OpenAI's application programming interface (API).

OpenAI has released three versions of its model this time: GPT-4.1, GPT-4.1 Mini, and GPT-4.1 Nano. Kevin Weil, the company's Chief Product Officer, stated during a live stream event that the new models outperform the current mainstream product GPT-4 in many aspects, and in some performance metrics, they even surpass the most powerful flagship model GPT-4.5.
OpenAI Unleashes a Major Move! GPT-4.1 Goes Live, Programming Power Fully Unlocked
In the authoritative SWE-Bench programming benchmark test, GPT-4.1 scored 55%, several percentage points higher than other OpenAI models. Weir said, "These new models are not only good at programming but also excel at precisely executing complex instructions, making them an excellent choice for developing 'agents'."

In recent months, the code generation capabilities of artificial intelligence models have significantly improved, driving the automation of software prototype development and the upgrading of "agent" capabilities. Competitors Anthropic and Google have both launched models that excel at code generation. Previously, there were widespread rumors in the industry that OpenAI would release GPT-4.1. According to informed sources, the company had anonymously tested the model under the code name "Alpha Quasar" on multiple evaluation platforms. Some Reddit users reported that this "anonymous" model "fixed all the incomplete code generated by other large language models."
The new model can analyze eight times the amount of code simultaneously compared to its predecessor, with significantly enhanced code optimization and vulnerability repair capabilities. Its ability to follow instructions has also been greatly improved, allowing users to achieve the expected results without repeatedly adjusting the instructions. In the demonstration, GPT-4.1 quickly built applications such as language learning flashcard programs based solely on natural language prompts.
"Developers place a high value on programming skills. We have focused on enhancing the model's ability to generate executable code," Michelle Pokrass, the head of the follow-up training team at OpenAI, revealed. "We have continuously improved the model's ability to follow multiple programming formats, enabling it to explore code libraries more deeply, reliably run unit tests, and write code that can be directly compiled."
OpenAI stated that GPT-4.1 is 40% faster than the most commonly used GPT-4 by developers, and the cost of querying input is reduced by 80%.

Varun Mohan, the CEO of the artificial intelligence programming tool Windsurf, stated in a live stream that during their internal benchmarking of GPT-4.1, they found that the new model's performance had improved by 60% compared to GPT-4.0. He disclosed that the new model exhibited significantly fewer erroneous behaviors and spent considerably less time mistakenly reading and editing irrelevant files.
Over the past few years, OpenAI has transformed public interest into a growing business by leveraging the frenzy sparked by the launch of its outstanding chatbot, ChatGPT, at the end of 2022. It has gradually built a commercial landscape encompassing multiple advanced models. Last week, the company's CEO, Sam Altman, revealed in a TED interview that OpenAI now has over 500 million weekly active users and is experiencing rapid growth.
Nowadays, OpenAI has launched a wide variety of models with different functions and pricing. In February this year, the company released its most powerful flagship model, GPT-4.5, but it is still in the "research preview" stage at present.
In addition, OpenAI has launched the o1 and o3 models with the ability to simulate reasoning. These models can break down problems into multiple parts to solve them, but they take longer to respond to users and have a higher usage cost.
The success of ChatGPT has led competitors to significantly increase their research and development investment in recent years, striving to catch up with OpenAI. According to a report released by Stanford University this month, models developed by companies such as Google and DeepSeek have reached performance levels comparable to those of OpenAI. The report also points out that companies like Anthropic, Meta, and the French startup Mistral are also close behind.
Oren Etzioni, an emeritus professor at the University of Washington, said that it is unlikely that a single model or a single company will monopolize the market in the future. "As costs decline, open source becomes more widespread, and specialized models for fields such as biology and chip design rise, the future will present a diversified development pattern," he said. Etzioni, who was formerly the head of the Allen Institute for AI (AI2), called on the industry to focus on the training costs and environmental impact of large models.
Facing profit pressure, OpenAI is accelerating its commercialization pace. In an interview in February this year, Brad Lightcap, the company's Chief Operating Officer, said that although the number of weekly active users has exceeded 400 million, a 30% increase compared to December 2024, OpenAI is still losing billions of dollars due to continuous heavy investment in research and development and infrastructure construction. In January this year, OpenAI announced that it would collaborate with SoftBank, Oracle, and MGX to advance the "Stargate" project, committing to invest 500 billion US dollars in building artificial intelligence data centers.
Recently, OpenAI has been continuously launching new features. Last week, OpenAI announced that ChatGPT would upgrade its "memory" function, enabling the chatbot to better remember and recall previous conversation content. At the end of March, OpenAI announced plans to release an open-source weight model this summer, allowing developers to download and modify it for free. The company will start testing within the next few weeks. Due to the fact that open-source weight models can be customized for different uses and have lower usage costs, they have been widely welcomed by researchers, developers, and start-ups.