Depositphotos
New family of OpenAI models includes GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano variants — all specialize in programming, have a context window of 1 million tokens (about 750 million words), and an output window of 32 thousand tokens, but are currently available only through the API.
The cost of using the models:
According to the company, GPT-4.1 is optimized for front-end coding, formatting, and tooling. On the SWE-bench Verified software development benchmark (52-54.6%), the model outperforms GPT-4o and GPT-4o mini, but lags slightly behind Gemini 2.5 Pro (63.8%) and Claude 3.7 Sonnet (62,3%). The company also clarifies that the reliability of the model decreases when processing a large number of tokens — from 84% for 8000 tokens to 50% for 1 million.
As a reminder, OpenAI plans to remove the GPT-4 model from ChatGPT on April 30, making GPT-4o «the natural successor to»; similarly, the previous version of GPT-4.5 will be discontinued in the API on July 14, as «GPT-4.1 offers improved or similar performance at much lower cost and latency than».
GPT-4o, which will become the base version of ChatGPT, last month updated with an option to generate images — and it became so popular that OpenAI had to limit requests and suspend access to free accounts to prevent «meltdowns» GPUs.
On April 4, OpenAI CEO Sam Altman announced on X that launch of GPT-5 has been postponed, so the model will appear only «in a few months», and not in May, as previously expected. In the coming days, we also expect the debut of the full versions of the o3 and o4 mini reasoning models — hints noticed in the latest ChatGPT weblog by engineer Tibor Blaho.