openai releases gpt model

openai releases gpt model

### Detailed Analysis of GPT-4.5: Features, Performance, and Technical Specifications (2025/03/06)

Summary:

This is a comprehensive news report from ITmedia AI+ detailing OpenAI's release of GPT-4.5, their largest language model to date. The article analyzes the model's technical specifications, performance improvements, and key features based on OpenAI's officially published system card and documentation.

Key Technical Features

  • Training Approach: GPT-4.5 focuses on unsupervised learning rather than chain-of-thought reasoning. This approach has led to improved world understanding and intuition without requiring extensive step-by-step reasoning.

  • Novel Training Method: The model uses a reverse knowledge distillation technique where data from smaller models (likely GPT-4o) is used to train a much larger model—contrary to traditional methods where larger models help train smaller ones.

  • Model Size: GPT-4.5 is described as "the largest LLM in OpenAI's history." While exact parameter counts are not disclosed, it's significantly larger than previous models, with CEO Sam Altman noting it's "massive and expensive" to run.

  • Training Data: The model uses publicly available data, proprietary data from partnerships, and custom datasets developed internally by OpenAI, all subjected to strict filtering processes.

Performance Improvements

  • Hallucination Reduction: Significant improvement in factual accuracy (78% on PersonQA benchmark compared to 28% for GPT-4o and 55% for o1) with hallucination rate reduced to 19% (from 52% in GPT-4o).

  • Multilingual Capability: Demonstrated improvements across various languages, including English (0.896 up from 0.887 in GPT-4o), Japanese (0.8693 up from 0.8349), and even low-resource languages like Swahili and Yoruba.

  • Conversation Quality: Enhanced understanding of user message nuances and more natural conversation flow.

  • Stereotype Resistance: Shows improved ability to provide responses that don't rely on stereotypes compared to previous models.

Safety and Risk Assessment

  • Safety Measures: Achieves 99% score in avoiding unsafe outputs, slightly better than GPT-4o (98%), but shows more caution with image inputs.

  • Risk Classification: OpenAI's safety advisory group rated GPT-4.5 as a "moderate" overall risk, with low risks in cybersecurity and model autonomy but moderate risks in CBRN (Chemical, Biological, Radiological, Nuclear) areas.

  • Jailbreak Resistance: Shows improved resistance to human-created jailbreak attempts but performs worse than GPT-4o and o1 on academic jailbreak benchmarks.

Practical Information

  • Context Window: 128,000 tokens, same as GPT-4o.

  • Maximum Output: 16,384 tokens, same as GPT-4o.

  • Knowledge Cutoff: October 2023, unchanged from previous models.

  • API Pricing: Extremely high at $75 per 1M input tokens and $150 per 1M output tokens—5x and 2.5x the cost of o1 respectively.

  • Availability: Initially available to Pro users immediately, with Plus and other paid plan users gaining access approximately one week later.

Future Outlook

The article notes that OpenAI is expected to announce GPT-5 in the coming months, which will reportedly eliminate the need for users to select specific models. GPT-4.5 appears to be an intermediate step in this development roadmap, showcasing OpenAI's continued focus on scaling up their language models while implementing new training methodologies.

Previousrefugees face port moresby evictionNextphysics breakthroughs summit updates