| GPT-4.1 | |
|---|---|
 | |
| Developer | OpenAI |
| Initial release | April 14, 2025; 7 months ago (2025-04-14) |
| Predecessor | |
| Successor | GPT-5 |
| Type | |
| License | Proprietary |
| Website | openai |
GPT-4.1 is alarge language model withinOpenAI'sGPT series. It was released on April 14, 2025. GPT-4.1 can be accessed through the OpenAI API or the OpenAI Developer Playground.[1][2][3] Three different models were simultaneously released: GPT-4.1,GPT-4.1 mini, andGPT-4.1 nano.[4] Since May 14, GPT-4.1 has been available for users subscribed to theChatGPT Plus and Pro plans, and GPT-4.1 mini that replacesGPT-4o mini is available for all ChatGPT users.[5]
All three models have acontext window of 1 million tokens and aknowledge cutoff of June 2024.[4]
The models were tested on numerousbenchmarks. Academic knowledge benchmarks included the 2024AIME,GPQA, andMMLU.[4] Coding benchmarks includedSWE-bench and SWE-Lancer.[4] Instruction following benchmarks included COLLIE and IFEval.[4] Vision benchmarks includedMMMU (answering questions about images),MathVista (solving vision-related mathematical tasks), and CharXiv (answering questions about charts from research papers).[4] Long-context benchmarks included two brand-new benchmarks invented by OpenAI: "multi-round coreference" (where the model has to find the i-th instance of something in a fake long conversation synthetically generated byGPT-4o)[6] and "Graphwalks" (forcing the model to simulatebreadth-first search).[4]
The models underwent more training regardingtool-calling, so the "OpenAI cookbook" recommends exclusively using the tools field when giving the model access to tools.[7] The models are also trained to follow instructions more literally, making the model more steerable.[7]
The Verge described GPT-4.1's release as "mark[ing] a pivot in the company's release schedule".[1]HackerNoon praised the model as "a HUGE win for developers", and stated that it challenged the advantages ofGemini 2.5 Pro's longer context window andClaude 3.7 Sonnet's strong reasoning capabilities.[8]Zvi Mowshowitz described GPT-4.1-mini as an "excellent practical model".[9] However, he criticized OpenAI for not doing enough safety testing, saying that he "hate[s] the precedent this sets".[9]
Two research teams - one led byOxford University researcher Owain Evans, the other based at theAI red-teaming startupSplxAI - independently found evidence that GPT-4.1 could be more misaligned thanGPT-4o.[10]
| Products |
| .svg){kind=logo} | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| People |
| ||||||||||||
| Joint ventures | |||||||||||||
| Related | |||||||||||||
| Concepts |
| ||||||
|---|---|---|---|---|---|---|---|
| Applications | |||||||
| Implementations |
| ||||||
| People |
| ||||||
| Architectures |
| ||||||
