Just days after announcing its plans to retire GPT-4 in ChatGPT, OpenAI on Monday launched a new set of flagship models named GPT-4.1.
The release, which The Verge anticipated in an article last week, included the standard version GPT-4.1 model, along with two smaller models – GPT-4.1 mini, and GPT-4.1 nano which OpenAI touts as its “smallest, fastest and cheapest model” till date.
All the models come with improved coding and instruction following capabilities, the company said. Notably, the new GPT-4.1 family boasts larger context windows handling up to 1 million context tokens including text, images and videos in prompts. By contrast, its predecessor, GPT-4o can support only up to 128,000 tokens.
“These models are better than GPT-4o on just about every dimension. They even meet or beat GPT-4.5 in a bunch of key ways,” said Kevin Weil, product lead at OpenAI, during a live demonstration.
The larger context windows allow the models to better use context with improved long-context understanding, said OpenAI, a capability considered particularly useful for developers who tend to use long and complicated prompts for programming use cases.
“We’ve trained GPT‑4.1 to be better able to pick out information from past messages in the conversation, allowing for more natural conversations,” the company said.
“We have been improving our model’s ability to write functional code…we’ve been working on making it follow diff formats, better explore repos, write unit tests and write code that compiles,” said Michelle Pokrass, post-training research lead.
The proof is in the pudding. According to SWE benchmarks, GPT-4.1 scores better in coding by 21.4% over GPT-4o and 26.6% over GPT-4.5. The score for instruction following is also up 10.5% over GPT-4o, according to OpenAI’s tests.
“Better instruction following makes existing applications more reliable, and enables new applications previously limited by poor reliability. Early testers noted that GPT‑4.1 can be more literal, so we recommend being explicit and specific in prompts,” the company said in the release note.
The models also have a more recent knowledge cutoff (up to June 2024), OpenAI said.
OpenAI positions GPT-4.1 as the “leading model” for coding. GPT-4.1 mini is recommended for lighter use cases, and GPT-4.1 nano for low-latency applications like autocompletion, classification and text extraction from long documents.
All three models are currently available to developers via OpenAI’s API, but not on ChatGPT.
Pricing for GPT-4.1 is at $2 per million input tokens and $8 per million output tokens. For GPT-4 mini, it is $0.40/million input tokens and $1.60/million output tokens, and for GPT-4.1 nano, $0.10/million input tokens and $0.40/million output tokens.
OpenAI said it will be turning off preview of GPT-4.5 in API in three months owing to the close and even superior performance shown by GPT-4.1.
