GPT-4.5

2025-02-27 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI releases GPT-4.5 as a research preview, its largest chat model, based on scaling up pretraining, with better intuition, fewer hallucinations and a higher “EQ”.

What changes

  • Largest model in the GPT lineup at this date, trained on Azure supercomputers, without explicit reasoning.
  • Broader world knowledge, better understanding of intent and a more natural style for writing, design and advice.
  • Access first for Pro subscribers, then Plus and Team the following week, then Enterprise and Edu.
  • API preview for developers on all paid tiers (Chat Completions, Assistants, Batch), with function calling, structured outputs and vision.
  • The vendor says it is evaluating whether to keep the model in the API; its retirement, announced on April 14, 2025, is set for July 14, 2025.

### Specifications - Context 128,000 tokens, maximum output 16,384 tokens, text and image input, text output. - Cutoff date: October 1, 2023. - API identifier: gpt-4.5-preview (snapshot 2025-02-27). - No voice mode, video or screen sharing in ChatGPT at launch. ### Pricing Prices per million tokens (USD), according to the API documentation. | | Input | Cached input | Output | | --- | --- | --- | --- | | gpt-4.5-preview | 75 | 37.50 | 150 | ### Benchmarks Scores published by the vendor (appendix of the announcement), compared with GPT-4o and o3-mini (high). | Benchmark | GPT-4.5 | GPT-4o | o3-mini (high) | | --- | --- | --- | --- | | GPQA | 71.4% | 53.6% | 79.7% | | AIME 2024 | 36.7% | 9.3% | 87.3% | | MMMLU | 85.1% | 81.5% | 81.1% | | MMMU | 74.4% | 69.1% | not disclosed | | SWE-bench Verified | 38.0% | 30.7% | 61.0% | ### Key takeaways GPT-4.5 targets writing and conversation tasks rather than reasoning, at a much higher cost than other OpenAI models. Its quick retirement from the API shows that its positioning remained experimental.

Links

Tags : LLM, OpenAI, GPT, Aperçu de recherche, Pré-entraînement