Altman consecutively released heavyweight news: GPT-5 will not only be freely available, but will also integrate multiple cutting-edge technologies. o3 and o4-mini will debut in a few weeks, and a mysterious open-source inference model is coming. However, on the other side, Meta's Llama 4 has been repeatedly delayed due to performance bottlenecks, making the AI race increasingly complex.
Altman finally admits: Training GPT-5 is too difficult, and we will delay its release by a few months.
But the good news is that o3 and o4-mini will be launched in a few weeks!
Although delayed from the original timeline, Altman's words once again heightened everyone's expectations for GPT-5.
There are many reasons, but the most exciting is that we discovered: We can make GPT-5 better than originally imagined! We also found that smoothly integrating everything is more difficult than expected. Additionally, we want to ensure we have sufficient computing power to handle unprecedented demand.
Altman also stated that OpenAI has truly improved the o3 model in many aspects and will definitely satisfy users.
Additionally, he replied to a netizen's comment, indicating that o3 Pro will be launched soon.
Just recently, someone discovered that a mysterious 4o model version accidentally appeared in the API and then disappeared. Perhaps this is the next iteration of 4o-mini?
Moreover, there was a similar leak on the 4th. OpenRouter had a mysterious model called "quasar-alpha", also claiming to be from OpenAI.
So, is OpenAI's 4o coming?
On Meta's side, Llama 4 has also been repeatedly reported to be delayed.
It seems DeepSeek's impact is indeed quite broad.
Altman: GPT-5 will be free, but you'll have to wait a bit
Previously, in a deep conversation with Silicon Valley's famous analyst Ben Thompson, Altman discussed the progress of GPT-5.
He also indicated that due to DeepSeek's influence, GPT-5 will consider making it free for users. This news instantly shocked the AI community.
Then came scattered preview messages.
According to Altman's X from February, here's the information about GPT-5:
GPT-4.5 (Orion) will be the last non-chain-of-thought model, about to be released
GPT-5 will unify all tools and functions, no longer releasing o3 separately
Free users can use GPT-5 unlimited, while paid users can access a "higher intelligence level" version
GPT-4.5 will be online in weeks, GPT-5 will be released in months
In other words, GPT-5 will become an all-powerful system that connects the o-series and GPT-series.
Since there will be no full-powered o3, GPT-5 will become an integrated system of multiple technologies in ChatGPT and API.
It can not only autonomously use tools but also understands when to think deeply, when to respond quickly, and can even handle various complex tasks.
OpenAI's purpose is self-evident - simplify internal model and product systems, making AI truly ready-to-use.
Altman empathetically said, "Like everyone else, we don't like the current model selection interface and hope to return to a simple and intuitive unified intelligent experience".
In summary, GPT-5 will integrate voice, Canvas, search, Deep Research, and other functions. The various functions launched on ChatGPT in the past few months will be unified in GPT-5.
One of OpenAI's primary goals is to unify their models and create a system that can integrate all our tools. These systems can determine when long-term thinking is needed, when it's not, and can be effective in various tasks.
For GPT-5's repeated delays, some netizens created this timeline to mock.
However, given that GPT-5 will be free, many people say it's worth waiting for.
It can be foreseen that free strategies for large models will become a major trend in the future.
Small and medium-sized manufacturers are likely to be gradually marginalized in fierce competition, partly due to high computing power costs and being squeezed out by big companies and unable to retain users.
An open-source model is coming, is it GPT-5?
The reason this news is so exciting is that just a few days ago, Altman solemnly announced that OpenAI will open-source a powerful inference model in the next few months.
This is the first open-source model from OpenAI since GPT-2. Since the timeline is in the next few months, it also implies a possibility: Will GPT-5 be open-sourced?
Previously, COO Brad Lightcap wrote a long article explaining why the team ultimately chose to open-source and why now.
The answer is actually simple: Developers, enterprises, and government customers made such a demand.
OpenAI serves millions of developers daily - many rely on a combination of proprietary and open-source models to build AI products.
Since 2020, they have provided cutting-edge models through API and released models like GPT-2 and Whisper to the open-source community.
As models improve, more people want to run them in various scenarios.
Through conversations with startups and developers, OpenAI clearly recognized the importance of supporting diverse needs, such as custom fine-tuning for specific tasks, more adjustable latency, running locally, or deployments requiring complete data control.
Of course, OpenAI will continue to provide cutting-edge models through API and ChatGPT, but API alone cannot fully satisfy many scenarios where developers want to build at their desired location or method.
Therefore, their goal in launching this open-source model is precisely to address this issue: expanding developers' access to powerful AI while maintaining high standards of safety and responsible deployment.
OpenAI API researcher Steven Heidel said, "This model can run on consumer-grade hardware".
Many netizens speculate that o1-mini open source is the most likely.
In any case, the free GPT-5 will arrive in a few months, and within a few months, OpenAI will open source a powerful reasoning model - two heavyweight developments that have the AI community eagerly anticipating.
Llama 4 Repeatedly Delayed
Contrary to Altman's constant leaks about release times, Meta has remained low-key.
It's worth noting that their Llama series models were once the kings of open-source models.
However, on the 4th, according to The Information, Meta plans to release its latest large language model, Llama 4, later this month.
Although Meta is going all out, hoping to take the lead in the AI race, the results seem less than ideal. They have already postponed it at least twice before.
The report quotes two informed sources saying that Meta might even delay the release of Llama 4 again.
The report states that one reason for Llama 4's delay is that it did not meet expectations in technical benchmarks, particularly performing poorly in reasoning and mathematical tasks.
Additionally, Meta is concerned that Llama 4's capabilities in human voice conversations are not as good as OpenAI's models.
This explains the repeated delays in Llama 4's release.
Meta absolutely cannot release a model that does not reach top performance. After all, they have invested too much in it.
According to reports, Meta plans to spend up to $65 billion this year expanding its AI infrastructure.
Meanwhile, DeepSeek's R1 has made people question whether developing top-tier AI models really requires billions of dollars.
The report also mentions that Llama 4 is expected to draw from some of DeepSeek's technologies, with at least one version being a Mixture of Experts (MOE) model. This approach trains different parts of the model for specific tasks, making them "experts" in their respective domains.
Additionally, Meta is considering first releasing Llama 4 through Meta AI, and then releasing it in open-source form.
References
https://twitter.com/sama/status/1908167621624856998
https://techcrunch.com/2025/04/04/openai-says-itll-release-o3-after-all-delays-gpt-5/
https://www.reuters.com/technology/artificial-intelligence/meta-nears-release-new-ai-model-llama-4-this-month-information-reports-2025-04-04/
This article is from the WeChat public account "New Intelligence", author: Aeneas Rhino, published with authorization by 36kr.
