GPT-4o

Informazioni su OpenAI

Fondato 2015 · San Francisco, USA

OpenAI was founded in December 2015 by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman as a non-profit AI research lab with a $1B pledged commitment. The lab restructured into a capped-profit company (OpenAI LP) in 2019 to attract capital from Microsoft, which has invested over $13 billion. OpenAI's most-cited papers include 'Improving Language Understanding by Generative Pre-Training' (GPT-1, 2018), 'Language Models are Few-Shot Learners' (GPT-3, 2020), the GPT-4 Technical Report (2023) and the InstructGPT/RLHF paper (2022). The ChatGPT consumer product, launched in November 2022, reached 100M weekly active users faster than any consumer software in history. Product milestones include GPT-3.5, GPT-4 (March 2023), GPT-4 Turbo, GPT-4o (May 2024), the o1/o3 reasoning family (late 2024/2025), GPT-4.1 (April 2025) and the Sora video model. Sam Altman is CEO. The company's last reported valuation in 2025 exceeded $300 billion.

Architettura

Decoder-only Transformer (natively multimodal)

GPT-4o (the 'o' stands for 'omni') was released in May 2024 as OpenAI's first natively multimodal flagship: a single neural network trained end-to-end on text, audio and images, replacing the earlier cascade of separate ASR, LLM and TTS pipelines used by ChatGPT Voice. Inputs and outputs can be any combination of text, audio and image tokens, and the model emits audio with a median response latency of 320ms for spoken conversation. Architecturally it is a decoder-only Transformer with a unified tokenizer that mixes BPE text tokens, image patches and discretised audio frames. Pretraining used a multi-trillion-token web-scale corpus filtered for quality, plus large image-text pair datasets and licensed audio. Post-training applied RLHF with human raters, model-graded rewards and red-teaming under OpenAI's Preparedness Framework. GPT-4o introduced a new tokenizer with substantially better compression for non-English scripts, reducing token counts by 1.1x-4.4x across languages such as Hindi, Arabic, Korean and Tamil. The context window is 128K tokens with 16K maximum output. Function calling, JSON mode, vision input and Structured Outputs (with strict schema adherence) are first-class. The model card discloses Preparedness Framework evaluations against CBRN, cyber and persuasion risks.

Parametri

Undisclosed (estimated in the hundreds of billions, dense)

Contesto

128.000 token

Capacità

• Natively multimodal: text, image and audio in and out
• Real-time voice conversation with sub-second latency
• 128K context window with 16K max output
• Improved non-English tokenizer (up to 4.4x fewer tokens for some languages)
• Vision: charts, diagrams, screenshots, OCR
• Function calling and parallel tool calls
• Structured Outputs with strict JSON schema
• Strong coding performance with file editing via tools
• Singing, emotional speech and laughter in audio mode
• Multilingual fluency across 50+ languages
• Best for: real-time voice agents, multimodal assistants, general-purpose chat, vision tasks.

Addestramento e licenza

Trained on a multi-trillion-token mixture of web text, code, books, licensed third-party data and large image-text and audio datasets. Knowledge cutoff is October 2023. OpenAI does not disclose exact dataset composition.

Licenza: Proprietary, available via OpenAI API and Azure OpenAI Service. Commercial use allowed under OpenAI Terms of Use.

Test di sicurezza: Evaluated under OpenAI's Preparedness Framework across CBRN, cyber, persuasion and model autonomy risks. External red-teamers included over 100 experts in chemistry, biology, security and misinformation.