OpenAI announced GPT-4.5 on February 27, 2025, as a research preview of what it called its largest and best model for chat, opening it first to ChatGPT Pro subscribers and to developers on paid API tiers. The model did not reason step by step before answering, and OpenAI said its size made it expensive to run. The API version was set to end on July 14, 2025, and GPT-4.5 was scheduled to leave ChatGPT on June 27, 2026.
Key takeaways
- The announcement is dated February 27, 2025. Pro users got access that day, Plus and Team users the following week, and Enterprise and Edu users the week after.
- On OpenAI's own SimpleQA test, GPT-4.5 scored 62.5% accuracy with a 37.1% hallucination rate, against 38.2% and 61.8% for GPT-4o.
- OpenAI announced the API wind-down on April 14, 2025. The reported price was USD 75 per million input tokens and USD 150 per million output tokens.
- GPT-4.5 was scheduled to leave ChatGPT on June 27, 2026, after a notice dated May 28, 2026.
What the February 2025 announcement said
OpenAI's post described GPT-4.5 as a step forward in scaling up pre-training and post-training, trained on Microsoft Azure AI supercomputers. It separated two lines of work: models that gain from more data and compute without a visible chain of thought, such as GPT-4 and GPT-4.5, and reasoning models such as OpenAI o1 and o3-mini. GPT-4.5 belonged to the first group. The company said it expected the model to hallucinate less, follow user intent better, and show greater "EQ," its shorthand for reading tone and implied expectations in conversation.
Availability was staged. ChatGPT Pro users could pick the model on web, mobile, and desktop from the announcement day. In ChatGPT it supported search, file and image uploads, and canvas, but not Voice Mode, video, or screensharing. Developers on all paid usage tiers could preview it in the Chat Completions, Assistants, and Batch APIs, with function calling, Structured Outputs, streaming, system messages, and image inputs.
The published numbers, side by side
Every figure below comes from OpenAI's own post. OpenAI ran the tests, and the post does not cite an independent replication.
| Test | GPT-4.5 | GPT-4o | Reasoning model shown |
|---|---|---|---|
| SimpleQA accuracy (higher is better) | 62.5% | 38.2% | o1: 47%; o3-mini: 15% |
| SimpleQA hallucination rate (lower is better) | 37.1% | 61.8% | o1: 44%; o3-mini: 80.3% |
| GPQA (science questions) | 71.4% | 53.6% | o3-mini (high): 79.7% |
| AIME 2024 (competition math) | 36.7% | 9.3% | o3-mini (high): 87.3% |
| SWE-Bench Verified (coding) | 38.0% | 30.7% | o3-mini (high): 61.0% |
GPT-4.5 beat GPT-4o on every row, while the reasoning models led on math and coding. OpenAI itself made that point, saying it expected pre-training and reasoning to complement each other in later models. The SimpleQA set contains short, hard factual questions, and a 37.1% hallucination rate means the model answered incorrectly on that share of them. It says little about everyday use. In side-by-side comparisons, human testers preferred GPT-4.5 over GPT-4o in between 56.8% and 63.2% of queries, depending on the category. The coding rows carry a note that the numbers show best internal performance.
Cost, the API, and the April 2025 wind-down
The launch post called GPT-4.5 a very large and compute-intensive model that was more expensive than GPT-4o and not a replacement for it. OpenAI added that it was evaluating whether to keep serving it in the API long term. On April 14, 2025, News coverage reported that OpenAI would end API access on July 14, 2025, and pointed developers to GPT-4.1, which launched the same day. An OpenAI spokesperson said GPT-4.1 offered similar or better performance in key areas at a much lower cost, and that GPT-4.5 was being deprecated to prioritize building future models.
News coverage listed the price as USD 75 per million input tokens and USD 150 per million output tokens, which it described as among the highest in OpenAI's catalog. At that point GPT-4.5 stayed available in ChatGPT as a research preview for paying customers.
What the system card reported
OpenAI published a system card on the same day as the launch. It said pre-deployment testing found no significant increase in safety risk compared with existing models. Under OpenAI's Preparedness Framework, the company's Safety Advisory Group rated GPT-4.5 as overall medium risk, with medium for chemical, biological, radiological, and nuclear (CBRN) topics and for persuasion, and low for cybersecurity and model autonomy. These are the developer's own classifications, published with its methods in the card.
Retirement from ChatGPT in 2026
OpenAI's ChatGPT release notes record the end of the model in two steps. On May 28, 2026, the company announced that GPT-4.5 would be retired after a 30-day sunset period, together with OpenAI o3, which was scheduled to go on August 26, 2026 after a 90-day period. Reports of the release notes say GPT-4.5 leaves ChatGPT on June 27, 2026 and is then no longer available in ChatGPT, including in custom GPTs. Existing conversations that used it can continue with GPT-5.5. OpenAI described the change as applying to ChatGPT only, with no change to the API. The stated reason was retiring older models with limited usage so the company can serve newer ones.
A checklist for reading any model launch
- Who ran the benchmark, and was the comparison model the closest rival or an older one?
- Does the post state a price, or only that one will follow?
- Is the release a preview, and does the company say how long it will stay available?
- Is there a system card or similar document with methods and limits?
- What happens to the previous model, and on what date?
FAQ
When was GPT-4.5 announced?
OpenAI announced it on February 27, 2025, as a research preview. ChatGPT Pro users and developers on paid API tiers received it first.
Is GPT-4.5 still available?
Not in ChatGPT. OpenAI scheduled its removal there for June 27, 2026. The API preview was set to end on July 14, 2025, according to the April 2025 announcement.
Did GPT-4.5 reason before answering?
No. OpenAI said the model does not think before it responds, which is what separates it from the o-series reasoning models.
How much did GPT-4.5 cost in the API?
News coverage reported USD 75 per million input tokens and USD 150 per million output tokens when the wind-down was announced in April 2025.
Related reading
- The earlier GPT-3 language model
- How a GPT chat bot generates language
- A timeline of AI history
- What a learning rate does in training
More in AI news.
Last verified: October 2026
Sources
- OpenAI, "Introducing GPT-4.5," February 27, 2025
- OpenAI, "GPT-4.5 System Card," February 27, 2025
- "OpenAI plans to wind down GPT-4.5, its largest-ever AI model, in its API," April 14, 2025
- OpenAI, ChatGPT release notes, entries of May 28, 2026 and June 2026
- OpenAI Help Center, "Legacy model access for Enterprise and Edu users," 2026