Will GPT-5 be released within two weeks and the exclusive details of its internal testing leak? GPT-6 may have started training, and Altman revealed a million GPUs.
Will GPT-5 be released within the next two weeks?
Just today, there's new breaking news about GPT-5.
Yuchen Jin, the co-founder of Hyperbolic who is well-informed, claims that he has learned some inside information.
GPT-5 is not a single model but a system composed of multiple models.
It includes a "router" that can switch between inference models, non-inference models, and tool-using models. That's why Altman said OpenAI will revise the model naming: in the future, prompts will be automatically routed to the most suitable model.
GPT-6 is already in training.
Another well-informed person has also confirmed this at the same time.
Actually, the coming of GPT-5 is not really new news.
When OpenAI researcher Alexander Wei announced on Saturday that the new model won the IMO gold medal, he had already previewed that GPT-5 would be released soon, but it's not the model that won the IMO gold medal.
Moreover, just yesterday, the words "GPT-5-reasoning-alpha-2025-07-13" appeared in the open-source code of a third-party institution.
The GPT-5 model was leaked in the biological benchmark test.
And this morning, Altman posted again, saying that by the end of this year, OpenAI will add more than 1 million GPUs, which obviously indicates that they are preparing more computing power for the new model.
Will GPT-5 be a game-changer or just a non-leap?
It's certain that GPT-5 will be born no later than September.
A few days ago, the mysterious model o3-Alpha was launched, but was removed from the public benchmark test by OpenAI just 12 hours later.
This may indicate that the official version is about to be released.
Historical data shows that when OpenAI tested confidential models such as "Optimus Alpha" and "Quasar Alpha", the official version of Quasar was released 11 days later, and Optimus Alpha was officially announced only 4 days later.
Regarding the upcoming GPT-5, some people are optimistic, while others are pessimistic.
OpenAI replaced the o3 model with gpt‑5 reasoning‑alpha.
For example, Ethan Mollick, a professor at the Wharton School, said that even if GPT-5 can only automatically switch between o3 and 4o, it will change most people's views on AI.
However, many people judge from various details that GPT-5 is likely to be a router.
For example, Kevin Weil, the CPO of OpenAI, revealed some clues about GPT-5 in February this year.
If GPT-5 is really just a router, the public will obviously be disappointed: it will obviously not have much improvement in basic intelligence, and we can only wait for Gemini 3 or Claude Sonnet 5.
Many people think that even if OpenAI really releases GPT-5, there won't be a significant improvement in the model's ability unless there are better tools or some ingenious methods to use RL to improve performance.
In short, many people who are looking forward to GPT-5 may be disappointed!
However, some people say that a router represents reliability and specialization. The reason why Erika is effective is that it distributes different types of overhead to the corresponding logical paths.
The underlying general intelligence may seem amazing in benchmark tests, but what can really be launched and scaled up is actually a specialized routing system.
In short, don't underestimate this kind of innovation. Sometimes, a seemingly ordinary architectural design may actually outperform a breakthrough model.
Altman breaks his silence, GPT-6 enters the endgame
However, the next-generation model, GPT-6, may start the endgame.
Some time ago, Altman shared his insights on the future development of AI again in a 20-minute interview with Sarah Guo, the founder of Conviction.
Altman said that the coding agent Codex released by OpenAI made him deeply feel the atmosphere of AGI.
Codex can not only handle complex tasks autonomously but also connect to GitHub and read internal documents, showing amazing capabilities.
Even, he predicted that AI agents may be like interns who work for a few hours today, but will evolve into senior engineers who work for days in the future.
Moreover, it will eventually become an "AI scientist" capable of discovering new knowledge - this will be an important moment for the whole world.
The host also asked, in the next-generation model, what "emergent behaviors" have you observed that will change the operation mode, product construction ideas, and the operation mode of OpenAI?
In response, Altman said very firmly that the models in the next 1 - 2 years will be very amazing, just like the major leap from GPT-3 to GPT-4.
As for what enterprises can do, it is to directly hand over the most difficult problems to the next-generation model.
Suppose a chip design company can let the LLM design a better chip. A biotech company aiming to cure a certain disease can also throw the problem to the AI.
Altman said that such a future is just around the corner.
As mentioned before, the LLM can understand any context, connect to every tool and every system, then conduct extremely excellent and intensive reasoning, and feedback high-quality answers.
Most importantly, they also have sufficient robustness and autonomy, and can be fully trusted to handle work.
Altman said excitedly again that he never thought this day would come so soon before, but now, it really feels very close.
He also proposed a Platonic ideal, a very small model with superhuman reasoning ability that can run at an incredibly fast speed, has a 1-trillion-token context, and can access all tools.
In this way, what the problem is no longer matters, and whether the model is pre-loaded with knowledge or a database also becomes unimportant.
People can regard it as a "reasoning engine" and just throw all possible contexts and relevant tools in an enterprise or a person's life to it.
Altman said that what people can achieve with it is quite amazing, and he thinks we are moving in this direction.
When asked what he would do if he had a thousand times more computing resources, Altman said he would let the AI research how to build a better model, and then ask the more powerful model how to use the resources.
At the same time, increasing the computing resources during testing can significantly improve the model's performance, especially when solving high-value problems.
Ten people at OpenAI rejected a $300 million offer
Meanwhile, a report from foreign media WSJ revealed a lot of inside information.
For example, at least ten employees at OpenAI rejected a $300 million offer from Zuckerberg.
Among those who rejected, there are people we are familiar with, such as Mark Chen, the chief research officer of OpenAI, and Noam Brown, the father of Texas Hold'em.
The news said that this spring, Zuckerberg had a simple meeting with Mark Chen and asked him for advice on how to improve Meta's generative AI team.
Unexpectedly, Mark Chen's words - increasing investment in talent - made Zuckerberg start