OpenAI released a new family of models for developers while revealing that it will no longer support GPT-4.5, which it released two months ago.
On April 14, the vendor launched three new API models: GPT-4.1, GPT-4.1 mini and GPT-4.1 nano.
According to OpenAI, the models perform better than GPT-4o and GPT-4o mini. They also have a larger context window and can support up to 1 million context tokens. Tokens refer to the unit of text that a model uses to process information.
The models were trained on information up to June 2024, OpenAI said.
The GPT-4.1 models are particularly effective for coding. They can edit large files across different formats.
The generative AI vendor said GPT-4.1 nano is its fastest and cheapest model for low-latency demand and is ideal for classification and autocompletion.
Besides being good at coding and having a long context window, the GPT-4.1 family understands images, surpassing GPT-4o on image benchmarks, according to OpenAI.
The introduction of GPT-4.1 comes about two months after OpenAI released a preview of GPT-4.5. With the launch of the GPT-4.1 family, OpenAI revealed that it will turn off the GPT-4.5 preview in the API by July 14 so that developers can't use it.
Direction and challenges
The launch of GPT-4.1 and the deprecation, or removal, of GPT-4.5 shows the direction OpenAI is taking and some of the challenges it is facing with its models.
Coding, or software engineering as a domain, has evolved as one of the domains where they're starting to see some real market traction, real revenue.
Arun ChandrasekaranAnalyst, Gartner
Like other model providers, OpenAI is moving away from using its models solely for generic purposes and toward using them for specific purposes, like coding.
"Coding, or software engineering as a domain, has evolved as one of the domains where they're starting to see some real market traction, real revenue," said Arun Chandrasekaran, an analyst at Gartner.
OpenAI is not the only generative AI vendor pursuing coding. In February, Anthropic introduced Claude Code, an agentic coding tool. Similarly, Google introduced Gemini Code Assist.
Another notable aspect of OpenAI's launch is the release of the vendor's first nano model.
"It kind of also tells you that perhaps the smaller models now have much better performance that allows them to suffice some of the more basic task requirements," said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget.
The nano model is also an example of OpenAI providing customers with what they want: a low-latency and low-cost model. Nano costs $0.10 per million tokens for input and $0.40 per million tokens for output. Comparatively, the other GPT-4.1 models' costs range from $0.40 to $2 per million input tokens and $1.60 to $8 per million output tokens.
Despite GPT-4.1 nano's lower cost, it is still not as cost-effective as Gemini Nano. Gemini Nano is free through Gemini Code Assist.
"As the use cases become more real-time, close to real-time, customers want low-latency performance," Chandrasekaran said. "Of course, the cost is becoming an increasing factor."
While OpenAI is trying to be more application-centered with its release and provide a lower-cost model, it might also use GPT-4.1 to address some challenges. Chandrasekaran said GPT-4.1 will only be available in the API, which could mean that the AI vendor is facing computing constraints.
"They're trying to optimize services where they can make more direct revenue using this model," he said. "It's an interesting juggling act that OpenAI is trying to do with the newer models in terms of just what is available as an API versus what is available within the application, and how quickly they seem to be deprecating some of these models."
Chandrasekaran added that for OpenAI to deprecate a model two months after introducing it to enterprises is confusing and not helpful because many enterprises can't afford to deprecate models so quickly.
Enthusiasm about agentic AI
However, the new family of models is the next step in the practical application of AI, such as coding and instruction-following, which will lead enterprises to build more intelligent systems and agentic applications, OpenAI said.
These advancements in the model market are what is creating the enthusiasm about what could be possible with agentic AI, Su said.
"We are edging toward a future of agentic AI that is exciting," he said. "We will see AI agents that are robust and that can do stuff that humans can't do. That is fundamentally what excites the industry. It's less about foundation models, but what they can build on top of the foundation models."
Another foundation model provider that is branching out is Cohere with its Command model. On April 15, Cohere introduced Embed 4, a multimodal model that enterprises can use to add search and retrieval capabilities to AI applications. This is essential for enterprises building AI assistants and agents, Cohere said.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
