GPT-4o Mini

Sobre OpenAI

Fundado em 2015 · San Francisco, USA

OpenAI was founded in December 2015 as a non-profit AI research lab by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman. In 2019 it transitioned to a capped-profit structure (OpenAI LP) to take Microsoft investment, now totalling over $13 billion. Foundational publications include the GPT series papers (GPT-1 through GPT-4), the InstructGPT/RLHF paper (2022) and the GPT-4o System Card (2024). The company ships ChatGPT (launched November 2022), the GPT API, the o-series reasoning models, Sora for video, DALL-E for images and Whisper for speech recognition. Sam Altman remains CEO; Greg Brockman is President. OpenAI's 2025 valuation exceeded $300 billion and the company employs more than 1,500 staff in San Francisco, London, Tokyo, Dublin and other offices. The mission statement focuses on building safe and beneficial AGI for all of humanity.

Arquitetura

Decoder-only Transformer (small natively multimodal)

GPT-4o mini was announced in July 2024 as a low-cost, fast replacement for GPT-3.5 Turbo. It is a small natively multimodal Transformer derived from the same training stack as GPT-4o, supporting text and vision input with text output. The model was trained on a multi-trillion-token web-scale corpus including code, books, licensed text and image-text pairs, with a knowledge cutoff of October 2023. Post-training combined supervised fine-tuning with RLHF and direct preference optimisation on smaller curated datasets. OpenAI applied an 'instruction hierarchy' training method to better resist jailbreaks and prompt injections by distinguishing system, developer and user instructions during alignment. At launch GPT-4o mini scored 82% on MMLU and outperformed GPT-3.5 Turbo on the Chatbot Arena leaderboard while costing approximately 60% less. It supports the full GPT-4o feature set including function calling, parallel tool calls, JSON mode, Structured Outputs and vision input. The model is the default backbone for ChatGPT Free as of mid-2024 and is widely used as a router and cheap tool-calling layer in agentic systems. Fine-tuning was opened to developers in late 2024 with both supervised fine-tuning and reinforcement fine-tuning options on the API.

Parâmetros

Undisclosed (estimated ~8B-20B parameters dense)

Contexto

128.000 tokens

Capacidades

• Very low cost per token (around $0.15 input / $0.60 output per 1M tokens at launch)
• 128K context window with 16K max output
• Vision input for images and PDFs
• Function calling and parallel tool calls
• Structured Outputs with strict JSON schema
• 82% MMLU at launch (better than GPT-3.5 Turbo)
• Trained with instruction hierarchy to resist prompt injection
• Supervised and reinforcement fine-tuning available
• Fast time-to-first-token for chat workloads
• Multilingual coverage but optimised for English
• Best for: cheap chatbots, classification, large-scale data labeling, agent routers, customer support.

Treinamento & licença

Pretrained on OpenAI's curated multi-trillion-token mixture of web text, code, books and image-text pairs, with a knowledge cutoff of October 2023. Post-training uses supervised fine-tuning, RLHF and the instruction-hierarchy alignment objective.

Licença: Proprietary, accessible via OpenAI API and Azure OpenAI Service. Commercial use permitted under OpenAI Terms.

Testes de segurança: Evaluated under OpenAI's Preparedness Framework and System Card process; smaller than GPT-4o so deemed low risk for CBRN uplift but tested against persuasion, cyber and prompt-injection threats.