Originally published March 3, 2025. Updated October 15, 2026. This article preserves the context of GPT-4.5’s February 2025 launch and updates its availability. It is historical reporting, not a current access guide.
Current status: OpenAI shut down gpt-4.5-preview in the API on July 14, 2025, after announcing deprecation on April 14. Separately, OpenAI’s current ChatGPT release notes state that GPT-4.5 became unavailable in ChatGPT, including custom GPTs, on June 26, 2026. These are different product timelines. Sources: OpenAI API deprecations and ChatGPT release notes, June 26, 2026.
What OpenAI launched in February 2025
OpenAI introduced GPT-4.5 as a research preview on February 27, 2025. Its emphasis was larger-scale unsupervised learning: improving knowledge, pattern recognition and conversational responses through pre-training. OpenAI distinguished this approach from the deliberate reasoning associated with o1 and o3-mini. It said GPT-4.5 was trained on Microsoft Azure AI supercomputers. See the original announcement.
What the benchmark results actually showed
One verifiable historical result is 62.5% on SimpleQA for gpt-4.5-preview-2025-02-27 in OpenAI’s simple-evals results, using the listed assistant prompt. The repository warns that prompting affects results and stopped updating new-model benchmark results in July 2025. Its table is a historical reference, not a current leaderboard.
That score is evidence about a particular factual-question evaluation. It does not mean that 62.5% of every user’s answers will be correct. Nor can subtracting the score from 100 establish a universal hallucination rate: errors, abstentions, task definitions and grading rules matter.
GPT-4.5 was not ahead on every launch benchmark. OpenAI’s announcement reported the following AIME 2024 math results:
| Model | Score |
|---|---|
| GPT-4.5 | 36.7% |
| GPT-4o | 9.3% |
| OpenAI o3-mini (high) | 87.3% |
The o3-mini comparison explicitly used its high setting. Treat these as provider-reported results under the published evaluation setup, not a guarantee about a different prompt, tool configuration or production task. A model can improve factual recall while still being a weaker choice on a particular reasoning test.
Conversation quality and practical applications
For a writer, the useful distinction is between an answer that sounds natural and one that preserves the intended meaning. For research, it is between fluent explanation and claims that can be traced to evidence. For coding, it is between plausible code and code that passes relevant tests. Those distinctions remain useful when reading older model announcements.
What the safety report did and did not establish
The February 27, 2025 system card describes supervised fine-tuning, reinforcement learning from human feedback and additional supervision techniques. OpenAI reported no significant increase in safety risk compared with existing models in its pre-deployment evaluations.
That is narrower than calling GPT-4.5 the safest model or saying testing guaranteed unbiased, harmless answers. The report describes remaining challenges involving hallucinations, jailbreaks and refusals. It also notes that production results can vary with system prompts, parameters and updates. Safety evaluation documents should be read as evidence of specific tests and limitations, not as certification that every output is trustworthy.
Historical access versus availability today
At launch, ChatGPT Pro users received access first, with Plus, Team, Enterprise and Edu rollouts announced to follow. ChatGPT features included search, uploads and canvas; voice, video and screensharing were excluded at launch. Paid developers could use the Chat Completions, Assistants and Batch APIs. These are historical launch details.
The model reference identifies the alias gpt-4.5-preview and dated snapshot gpt-4.5-preview-2025-02-27, marked deprecated. An old documentation page remaining online does not establish that the model is still served. Use the API shutdown record and ChatGPT retirement notice above to interpret older tutorials.
For more on interpreting evaluation results, see why benchmark scores alone are not enough to choose an assistant.
The lasting takeaway
When revisiting GPT-4.5 coverage, keep three records separate: the launch announcement, the exact evaluation result and the product’s availability timeline. None substitutes for the others.
If an old workflow names GPT-4.5, first identify whether it used ChatGPT or an API model ID. Then list the behavior the workflow depended on, such as preserving a document’s meaning, extracting cited facts or returning valid structured data. Assess any replacement against those requirements using representative examples and human review. This is a practical migration approach, not a claim that we have tested a replacement or that a particular current model will reproduce GPT-4.5’s behavior.
