OpenAI’s new GPT-4.1 models can process a million tokens and solve coding problems better than ever

Openai has launched a new family of artificial intelligence models this morning that greatly improves coding capabilities while reducing costs, and directly responding to the increasing competition in the AI ​​market for institutions.

San Francisco IQ of San Francisco presented three PPT-4.1, GPT-4.1 Mini and GPT-4.1 Nano-all-available through its application programming interface. The new collection works better in software engineering tasks, tracks the instructions more accurately, and can process up to one million symbols of context, equivalent to about 750,000 words.

“GPT-4.1 performs exceptionally at a lower cost,” Kevin Whale, chief product official at Openai, said during the Monday announcement. “These models are better than GPT-4O at almost every dimension.”

Perhaps the most important prices for institutional customers: GPT-4.1 will cost 26 % than its predecessor, while the lightweight nano version becomes more likely in Openai at a price of only 12 cents per million icons.

https://www.youtube.com/watch?

How to target GPT-4.1 improvements for institutions

In an explicit interview with Venturebeat, Michel Boukerras stressed post -training research in Openai, that practical business applications prompted the development process.

“GPT-4.1 was trained with one goal: being useful for developers,” Pokrass told Venturebeat. “We have found GPT-4.1 much better to follow the types of instructions that institutions use in practice, making it much easier to publish ready-to-production applications.”

This focus is reflected on the tool in the standard results. On Swe-Bused, which measures software engineering capabilities, GPT-4.1 record 54.6 %-a significant improvement at 21.4 percentage on GPT-4O.

For companies that develop artificial intelligence agents who work independently on complex tasks, improvements in the following instructions are of special value. On the Scale, the GPT-4.1 38.3 %, outperform GPT-4O by 10.5 percent.

Why challenge the three -level model strategy of Openai such as competitors such as Google and Noteropic

The introduction of three premium models is treated with different price points, the diverse artificial intelligence market. The pioneer GPT-4.1 targets complex institutions applications, while Mini and NANO versions deal with speed use and cost efficiency of priorities.

“Not all tasks need most smart or higher capabilities,” Boukras told Venturebeat. “Nano will be a model for the backbone of use cases such as automatic completion, classification, data extraction or anything else as the speed is the highest concern.”

At the same time, Openai has announced plans to neglect the GPT-4.5 preview-its largest and largest model just two months ago-from its application programming interface by July 14. The company has put GPT-4.1 as a more cost-effective alternative that provides “improved or similar performance on many key capabilities at a much lower cost.”

This step allows Openai to restore computing resources while providing developers with a more efficient alternative to its most provided, which was priced at $ 75 per million input codes and $ 150 per million output symbols.

Real World Results: How Thompson Reuters, Carlel and Windsurf GPT-4.1 Invest

Many institution agents who tested models before launching significant improvements in their specific fields.

Tomson Reuters has witnessed a 17 % improvement in the accuracy of a multi-document review when using GPT-4.1 with legal AI assistant, Cocoonsel. This improvement is a special value for the complex legal workflow, which includes lengthy documents with accurate relationships between the sentences.

Carlyle’s financial company has reported 50 % better to extract granular financial statements from dense documents-a decisive ability to analyze investment and decision-making.

Varun Mohan, CEO of the Coding Tool Windsurf (formerly Codeium), joint detailed performance standards during the announcement.

“We have found that GPT-4.1 reduces the number of times that need to read unnecessary files by 40 % compared to other leading models, and it also modifies unnecessary files by 70 %,” Mohan said. “The model is also less amazing … GPT-4.1 is 50 % less than the other leading models.”

Millions context: What companies can do with 8x processing capacity more

All three models feature a context window of one million symbols-eight times the distinctive GPT-4O 128000. This expanded capacity allows models to process long documents or complete symbolic rules at once.

In a demonstration, Openai GPT-4.1 showed the NASA Service Register file analysis 450,000 of 1995, with an anomalous entry hidden in the depth of data. This capacity is of special value for tasks that include large data collections, such as code warehouses or corporate document groups.

However, Openai admits the deterioration of performance with very large inputs. In the internal Openai-MRCR test, the accuracy decreased from about 84 % with 8000 icons to 50 % with a million icons.

How to turn the AI ​​Enterprise Ai scene with Google, Anthropic and Openai for developers

This version comes as a competition at Space Ai Enterprise. Google recently launched Gueini 2.5 Pro with a window of a million million million, while Claude 3.7 Sonnet gained human beings towards companies looking for alternatives to Openai’s offers.

The Chinese company Deepseek Startup has also promoted its models recently, prompting additional pressure on Openai to maintain its leadership position.

“It was really great to see how improvements in the long context had been translated into better performance on specific heads such as legal analysis and extracting financial statements,” said Boukras. “We have found that it is important to test our models outside the academic standards and ensure that they perform well with companies and developers.”

By launching these models specifically through its application programming interface instead of ChatGPT, Openai indicates its commitment to developers and institutions agents. The company plans to gradually integrate features of GPT-4.1 in Chatgpt over time, but the primary focus on providing strong tools to create specialized applications remains.

To encourage further research in tall context processing, Openai issues two data evaluation groups: Openai-MRCR to test multi-cyclical basics capabilities and graphic graphics for complex thinking evaluation via long documents.

For decision makers at the Foundation, the GPT-4.1 family provides a more expensive and effective approach to the implementation of artificial intelligence. With institutions continuing to integrate artificial intelligence in their operations, these improvements in reliability, privacy and efficiency can accelerate adoption through industries that still weigh the costs of implementation against possible benefits.

While competitors are chasing larger and more expensive models, Openai’s strategic axis with GPT-4.1 indicates that the future of artificial intelligence may not belong to the largest models, but to the most efficient models. The real penetration may not be in the standards, but in bringing artificial intelligence at the level of the institution more than more companies than ever.