GPT-4.5
GPT-4.5, OpenAI's largest non-reasoning model, ships as a research preview at $75/$150 per million tokens, then is retired from the API within months.
Scaled pretraining and post-training for broader knowledge and "EQ", with no chain-of-thought reasoning; same 128K context and image input as GPT-4o. SimpleQA hallucination rate 37.1% versus 61.8% for GPT-4o. Too costly to serve, with API deprecation announced 2025-04-14 and shutdown 2025-07-14.
- Date
- Thursday, 27 February 2025
- Lab
- OpenAI
- Kind
- model
- Access
- research preview
- Price
- $75 input / $150 output per 1M tokens (2025-02-27)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SimpleQA hallucination rate (lower is better) | 37.1% GPT-4o 61.8%, o1 44%, o3-mini 80.3% | company |
| Aider polyglot | 45% vs Claude 3.7 Sonnet 60-65% and o3-mini 60.4% | independent |
| API price (input / output per 1M tokens) | $75 / $150 30x GPT-4o input price | company |
As an inference, it was the end of the pure pretraining-scale line, since GPT-5 later unified chat and reasoning models behind a router. OpenAI itself said at launch it was evaluating whether to keep serving GPT-4.5 in the API.
Sources
- openai.com/index/introducing-gpt-4-5/
- simonwillison.net/2025/Feb/27/introducing-gpt-45/
- developers.openai.com/api/docs/deprecations
This record was checked against its sources on 6 October 2026. How we check