For paid users.

The non-reasoning GPT-4.1 is available in the ChatGPT model list for Plus, Pro, and Team subscribers, OpenAI announced on X. Enterprise and Edu users will get access in the coming weeks.
The GPT-4.1 mini model will replace GPT-4o mini in ChatGPT. It will be available to all users.
OpenAI introduced GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano in April 2025. According to the company, they surpass GPT-4o and GPT-4o mini and are particularly good at writing code and following instructions. Initially, the models were available via API.

How the GPT-4.1 family models differ

A user asked all models in the lineup to create an HTML file for an animation of several rotating hexagons with balls.

Where to test GPT-4.1 and how much it costs

In OpenAI's API, GPT-4.1 costs $2 per 1 million input tokens and $8 per 1 million output tokens; GPT-4.1 mini — $0.40 and $1.6 respectively; GPT-4.1 nano — $0.10 and $0.40 respectively.
The developers of the code editor Cursor AI added the model to their service and opened free access — temporarily. A Cursor AI subscription costs from $20 per month.
Windsurf also allowed free testing of the model — until April 20, 2025. The minimum subscription cost is $15.

How the price and capabilities of GPT-4.1 and competitors compare

As independent researchers from Artificial Analysis write, GPT-4.1 is smarter and cheaper than GPT-4o. According to their tests, GPT-4.1 outperformed Llama 4 Maverick, Claude 3.7 Sonnet, and GPT-4o, and also matched the new V3 version from DeepSeek.
GPT-4.1 mini, according to their data, slightly surpasses GPT-4.1 in programming. GPT-4.1 nano is roughly equivalent to Llama 3.3 70B and Llama 4 Scout.

Model scores based on test results. Source: Artificial Analysis

Model scores based on test results. Source: Artificial Analysis

Graph of the price-to-capability ratio of models. Source: Artificial Analysis

Graph of the price-to-capability ratio of models. Source: Artificial Analysis

Artificial Analysis did not test reasoning models. In a general programming test by the code editing platform Aider, among all model types, the new reasoning Gemini 2.5 Pro took first place, while GPT-4.1 was in 13th place, one of the developers noted on X.
Those who tested the new OpenAI model added in response that they were not surprised by the result: they believe that Gemini 2.5 Pro writes code better — primarily thanks to long chains of reasoning.

What impressions developers are sharing

One developer tested how the model creates frontend code. Claude 3.7 Sonnet wrote twice as much code but, instead of selecting images, used gray placeholders. The user liked GPT-4.1's result more.

GPT 4.1 result for the prompt: Create a frontend for a movie streaming service in a single HTML file. Source: AiBattle

GPT 4.1 result for the prompt: Create a frontend for a movie streaming service in a single HTML file. Source: AiBattle

Claude 3.7 Sonnet result. Source: AiBattle

Claude 3.7 Sonnet result. Source: AiBattle

Another developer asked two models to create a note-taking app. He, on the contrary, rated Claude's result higher, and called GPT-4.1 and other models in the lineup more lazy.

Claude 3.7 result (left) and GPT-4.1. Prompt: An iOS journaling app that stores past entries as a list. Source: Josh Johnson

Claude 3.7 result (left) and GPT-4.1. Prompt: An iOS journaling app that stores past entries as a list. Source: Josh Johnson

A developer of educational apps for children noticed that GPT-4.1 reads fewer unnecessary files, makes fewer useless changes, and isn't as verbose.

GPT-4.1 created a "simulator" for learning about physical phenomena in a game format.

Source: Parul Pandey

Source: AiBattle

Another user tested how models create drawings in SVG format. Prompt: "Create a beautiful SVG with an image of the first three first-generation starter Pokémon in a single HTML file."

In March 2025, AI enthusiasts developed a humorous creativity test for neural networks, the Minecraft Benchmark. Users choose the best build out of two, without knowing which model made it. Based on these preferences, a model ranking is compiled.
Currently, the leader is Gemini 2.0 Pro. The developers haven't added GPT-4.1's position yet, but they have already shared the model's first generations.

Glass palaces by GPT-4.1 (right) and Gemini 2.5 Pro Experimental:

Earth "through the eyes" of GPT-4.1 (right) and GPT-4.5:

What to consider in prompts for the new GPT-4.1

OpenAI adapted GPT-4.1 for creating AI agents and working with long contexts. The developers published instructions for crafting prompts.
GPT-4.1 can follow instructions more precisely, avoiding the free interpretations that previous models might have allowed, so the main thing is to formulate the request clearly.
Developers claim that GPT-4.1 can handle even the maximum request length of 1 million tokens. But they advise writing structured instructions—both at the beginning and at the end of a long request.

#news #openai #chatgpt