GPT-4.1

O OpenAI

Založené 2015 · San Francisco, USA

OpenAI was founded in 2015 as a non-profit research lab by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman. In 2019 it transitioned into a capped-profit company to raise capital from Microsoft, which has invested over $13 billion. OpenAI is the publisher of the GPT series papers (GPT-1 in 2018, GPT-2 in 2019, GPT-3 in 2020, GPT-4 Technical Report in 2023), the InstructGPT paper that introduced RLHF for instruction following, and the GPT-4o System Card. Major products include ChatGPT (launched November 2022, with over 200M weekly active users), the OpenAI API, the o-series reasoning models, Sora for text-to-video and DALL-E for image generation. GPT-4.1 was released in April 2025 as an API-only model family (4.1, 4.1 mini and 4.1 nano) optimised for developers, with major improvements over GPT-4o on coding, instruction following and long-context tasks. The 2025 reported valuation exceeded $300 billion.

Architektúra

Decoder-only Transformer (developer-focused flagship)

GPT-4.1 launched in April 2025 as an API-only successor to GPT-4o aimed primarily at developers and agentic workloads. It is a decoder-only Transformer trained on an updated multi-trillion-token mixture of text, code and image-text pairs, with a knowledge cutoff of June 2024. The model family ships in three sizes (4.1, 4.1 mini, 4.1 nano) sharing the same training recipe but distilled to different capability/latency points. The headline change versus GPT-4o is a 1,047,576-token context window (~1M tokens), validated with new long-context evaluations such as OpenAI's MRCR (Multi-Round Coreference Resolution) and Graphwalks benchmarks. Post-training emphasised coding (SWE-bench Verified moved to 54.6% from 33% on GPT-4o), instruction following on edge-case formats, and reliable behaviour in long agent loops. OpenAI used reinforcement learning against verifiable rewards and large-scale synthetic data generation to push these axes. The model deprecates GPT-4.5 Preview from the API and is positioned as the production workhorse for ChatGPT Enterprise and Team deployments. Vision input is supported in all three sizes, function calling and Structured Outputs are first-class, and pricing is roughly 26% cheaper than GPT-4o at launch.

Parametre

Undisclosed (estimated multi-hundred billion parameters dense)

Kontext

1 047 576 tokenov

Schopnosti

• 1,047,576 token context window (~1M tokens)
• Major coding improvement: 54.6% on SWE-bench Verified at launch
• Strong instruction following on edge-case formats and negative constraints
• Reliable behaviour over long agent loops
• Vision input across all three model sizes (4.1, mini, nano)
• Function calling, parallel tool calls and Structured Outputs
• Knowledge cutoff June 2024 (later than GPT-4o)
• About 26% cheaper than GPT-4o at launch
• Diff-style file editing and applied-diff coding workflows
• Improved multilingual non-English benchmark performance
• Best for: coding agents, IDE copilots, large-document analysis, production assistants.

Tréning a licencia

Pretrained on a multi-trillion-token web-scale mixture of text, code repositories, books, licensed datasets and image-text pairs. Knowledge cutoff is June 2024. Post-training combined supervised fine-tuning, RLHF and reinforcement learning against verifiable rewards (especially for code).

Licencia: Proprietary commercial license via OpenAI API and Azure OpenAI Service. Not available through the ChatGPT consumer product.

Bezpečnostné testy: Evaluated under OpenAI's Preparedness Framework v2 across CBRN, cyber, persuasion and autonomy axes. External red-teamers and OpenAI's Safety Advisory Group reviewed before launch.