GPT-5.4

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 organisation; it transitioned to a 'capped-profit' structure in 2019. OpenAI authored the GPT lineage starting with GPT-1 (2018), GPT-2 (2019), GPT-3 (2020), the ChatGPT launch with GPT-3.5 (November 2022), GPT-4 (March 2023), GPT-4o (May 2024), the o-series reasoning models o1 (December 2024) and o3 (early 2025), then GPT-5 (mid-2025). The GPT-5.x series unified the reasoning o-series with the general-purpose GPT line: GPT-5.2 shipped December 2025, GPT-5.3-Codex February 2026, GPT-5.4 in March 2026 and the GPT-5.5 family later in 2026. OpenAI's investors include Microsoft (~$13B+), Khosla Ventures, Andreessen Horowitz, Thrive Capital and Sequoia, with total funding exceeding $60 billion and a 2026 valuation above $300 billion.

Architettura

Unified Transformer with integrated reasoning ('Thinking') tier

GPT-5.4 is OpenAI's frontier model as of March 2026 and marks the full convergence of the previously separate o-series reasoning models and general-purpose GPT line into a single unified architecture. The o1, o3, o3-pro and o4-mini reasoning models were retired from ChatGPT on February 13, 2026; their capabilities are now folded into GPT-5.x as a 'Thinking' tier that activates on demand. GPT-5.4 also absorbs the Codex line for agentic coding. Pretraining used a multi-trillion-token mixture of web text, code repositories, scientific papers, books and licensed datasets on OpenAI's most recent supercomputer cluster. Post-training combined supervised fine-tuning, RLHF and large-scale reinforcement learning against verifiable rewards on coding, math and tool-use trajectories. The model supports native text + image input and a 1.05M-token experimental context window (272K standard), 128K max output, integrated tool-use, function calling, structured outputs, and built-in 'Tool Search' for delegating to subagents. Safety training followed OpenAI's Preparedness Framework with red-teaming, capability evaluations and external assessments.

Parametri

Undisclosed (estimated multi-hundred billion parameters, likely sparse MoE)

Contesto

1.050.000 token

Capacità

• Unified GPT and o-series: integrated 'Thinking' tier activates on demand
• 1.05M-token experimental context window (272K standard, 128K max output)
• Native multimodal input: text and images in a single pass
• Top scores on SWE-Bench Pro and OSWorld-Verified at launch
• Native tool use, function calling, parallel tool calls and structured outputs
• Integrated 'Tool Search' for delegating to subagent models
• Absorbed Codex agentic coding capabilities
• Strong multilingual performance and instruction following
• Available via OpenAI API, Azure OpenAI, ChatGPT and Codex CLI
• Regional processing (data residency) endpoints available with 10% uplift
• Best for: high-context coding, agentic engineering, multimodal analysis, complex tool-use workflows.

Addestramento e licenza

Pretrained on a multi-trillion-token mixture of web text, code, scientific papers, books and licensed datasets. Post-training combines supervised fine-tuning, RLHF and large-scale reinforcement learning against verifiable rewards on coding, math and tool-use tasks. Knowledge cutoff approximately late 2025.

Licenza: Proprietary commercial license via OpenAI API and Azure OpenAI. Commercial use permitted under OpenAI's Usage Policies.

Test di sicurezza: Evaluated under OpenAI's Preparedness Framework with internal and external red-teaming, capability evaluations and a public Bug Bounty programme.