[New Intelligence Guidance] In the GPT-4 era, the curtain falls! It will be completely replaced by GPT-4o. Next week, we are also very likely to welcome new models such as GPT-4.1, o3 full version, and o4-mini. The history of large models is once again reaching a critical point.
Unexpectedly, GPT-4 has thus exited the historical stage.
On April 10th, OpenAI officially announced that it will remove the two-year-old GPT-4 from ChatGPT.
From April 30th, GPT-4 will be completely replaced by the current default model GPT-4o.
Of course, users who miss GPT-4 need not be sad, as it can still be used through the OpenAI API.
However, before GPT-4's retirement, people's attention will likely turn to ChatGPT's new models.
Now, various leaks are flying everywhere. It is said that GPT-4.1 (including GPT-4.1-mini, GPT-4.1-nano, and GPT-4.1), the o3 "reasoning" model announced last December, and the new o4-mini reasoning model are all coming.
Recalling March 14, 2023, when GPT-4 was officially released, it swept the globe, creating a massive wave, and within two years, it triggered a Cambrian explosion of global AI large models.
Its emergence opened a brand new era, witnessed a technological breakthrough, and paved the way for the subsequent prosperity of numerous OpenAI large models.
Now, this beacon is finally about to exit.
[The rest of the translation follows the same approach, maintaining the structure and translating all text while preserving the image tags and other HTML elements.]Among them, GPT-4.1, an improved version of the multimodal model GPT-4o, is expected to be released as early as next week.
Currently, AI engineer Tibor Blaho has discovered references to these models in the ChatGPT web code.
However, with GPT-4.5 already in the works, why is the next version GPT-4.1? This increasingly regressive naming convention is truly puzzling.
It seems somewhat out of place in OpenAI's large model lineup, with no clear positioning.
Some exclaimed: Seeing the evolution from GPT-1 to the o series and beyond demonstrates increasing complexity and capabilities. It's like observing the history of artificial intelligence evolution!
Regarding GPT-4.1, there are already some clues circulating.
For instance, some discovered that GPT-4.1 has already appeared as an option in ChatGPT's model list.
This means that the previously leaked rumors are true! This has made everyone very excited.
But some suspect that the user might have fabricated it by changing the URL from GPT-4o to GPT-4.1.
To prove their innocence, this user generously shared their conversation link with GPT-4.1.
Therefore, everyone confirmed that the user did not lie or forge screenshots, as checking the conversation link reveals that the model-slug is indeed GPT-4.1.
Soon, more and more people discovered they could also use GPT-4.1.
Additionally, some users speculate that Altman's "Quasars Alpha" refers to GPT-4.1.
To Prepare for o3 Release, Testing Time Significantly Shortened
The FT report also confirms that the full-featured o3 will likely be online again next week.
Insiders revealed that due to intense pressure from competitors like Meta, Google, and xAI, OpenAI had to accelerate model release speed to maintain its technological advantage and profit.
The model "assessment" that previously took months to test risks and performance has now been compressed to just a few days.
For example, the soon-to-be-retired GPT-4 had a 6-month assessment period. Some dangerous capabilities were only discovered two months into testing.
But now, the given time and resources are insufficient for comprehensive testing to identify and mitigate risks.
Previously, OpenAI promised to build custom models to assess potential abuse risks, such as whether they could enhance biological virus transmission capabilities.
However, this method is costly, requiring hiring external experts, creating specialized information datasets like virology, and "feeding" them to the model for fine-tuning, consuming significant engineering and computational resources.
Clearly, OpenAI doesn't want to invest too much effort in this—they only fine-tuned less capable older models and didn't apply it to more powerful, advanced models.
For instance, in the January safety and performance report for o3-mini, they only mentioned that GPT-4o would execute certain biological tasks after fine-tuning. As for o1 and o3-mini, OpenAI said nothing.
Another concerning issue is that safety tests target earlier "checkpoints" rather than the final publicly released models. After this, models continue to be updated to enhance performance and capabilities.
Facing these doubts, OpenAI responded that its assessment process has improved efficiency by introducing automated testing, thus shortening the required time.
"Although there's no universally accepted approach, the method OpenAI adopted is certainly optimal and has been transparently explained in the report."
Additionally, OpenAI stated that the checkpoints used for testing are "essentially consistent" with the final released version.
References:
https://x.com/whylifeis4/status/1910719046665884006
https://www.ft.com/content/8253b66e-ade7-4d1f-993b-2d0779c7e7d8
https://techcrunch.com/2025/04/11/openai-is-winding-down-its-gpt-4-ai-model-in-chatgpt/
Editor: Aeneas Good Tired
This article is from the WeChat public account "New Intelligence" and is published by 36kr with authorization.
