OpenAI appears to be preparing for a preview release of GPT-4.5. This development was discovered in ChatGPT’s Android app as an announced experimental option, although it is not yet accessible. The LLM will initially be available exclusively to Pro subscribers, according to the notification.
GPT-4.5 may become the successor to GPT-4o in ChatGPT’s free version, with much higher call limits and/or early access for Plus and Pro users. For now, however, GPT-4.5 has only made itself known via a notification in the mobile app. There is no way to use the LLM yet; pressing the notification does not accomplish anything.
It isn’t the first time the existence of GPT-4.5 appeared to be revealed. In late 2023, it appeared ChatGPT hallucinated this model name in a conversation. Other AI models sometimes misidentify themselves as well, such as DeepSeek, which thought it was ChatGPT for a moment.
Exclusively for Pro users
The new notification invites Pro users to test OpenAI’s “newest and largest model”. The Pro subscription, which costs $200 a month, offers fewer restrictions than the $20 Plus subscription or the free plan.
In addition to the introduction of GPT-4.5, OpenAI is working on improvements to its Web interface. AI researcher Tibor Blaho has discovered that features such as Deep Research, Canvas and the ability to upload images have been moved. This is part of a redesigned interface that allows users to switch between models and functions more easily.
For Deep Research mode, a new input option has been added. Users can add relevant sources, areas of interest and files for detailed reports on them. OpenAI has not yet announced when GPT-4.5 and the interface enhancements will actually be available.
Also read: OpenAI comes out with ‘deep research’; the answer to DeepSeek?
Competitors everywhere
OpenAI has already recently released o3-mini, an AI model that can reason while keeping its token spend relatively low. It is considerably cheaper than the earlier o1, but gets close to its performance in benchmarks. Still, OpenAI has more than enough competition in this field, including alternatives that are significantly more affordable. For example, Google has made available a Gemini Flash Thinking model. Earlier, DeepSeek R1 caused an uproar on the stock market among AI companies because it had similar functionality to OpenAI’s o1 but was based on a model (DeepSeek V3) that was said to have costed just $5 million to train. This made many suddenly doubt the massive capital spend AI companies are planning.
Anthropic also unveiled a long-awaited update to Claude this week with 3.7 Sonnet. In addition, the CLI tool Claude Code appeared.