GPT-5 Leaks Reveals Better Mathematics And Coding Capabilities

GPT-5 is able to provide a better answer even if it is only with a short prompt (photo: dock. openai)

The latest leaks reveal a number of details about the latest AI model from OpenAI, namely GPT-5, which is rumored to be bringing about a major increase in reasoning, mathematics, and coding capabilities. However, it is not yet clear when these features will be available on ChatGPT.

This information originally emerged from a post blog on GitHub which was accidentally published and later deleted. A Reddit user discovered the blog and shared details (via The Verge), including significant improvements offered by GPT-5 compared to previous models such as GPT-4, GPT-4.1, and GPT-4o.

According to the leak, the GPT-5 was able to provide a better answer even though it was only with a short prompt, showed a clearer mindset, and increased collaboration skills with users in various contexts.

The deleted but still accessible blog through the archive also revealed that the GPT-5 will be present in four different variants:

GPT-5: Standard variant for logic and multi-step task.

GPT-5-mini: For more cost-effective implementations.

GPT-5-nano: For a high-speed query response.

GPT-5-chat: Designed for multimodal conversation workflows, suitable for company needs.

The GPT-5 is also reported to have better agentic capabilities (agentic capabilities), although it is not explained in detail. OpenAI itself introduced an agent feature in ChatGPT last month, which allowed users to delegate tasks to run directly on the server instead of on the user's device.

While there has been no official announcement from OpenAI, the apparently very complete GitHub blog shows that the launch of the GPT-5 is likely very close.

In fact, OpenAI recently announced that it will hold a live stream today, Thursday at 10 am Pacific time (at 00.00 WIB Friday morning). Interestingly, in the announcement, the letter S in the word Livestream is replaced by a 5 number of strong clues to the launch of the GPT-5.

OpenAI co-founder Sam Altman previously discussed the experience of testing this new model on a podcast with Theo Von. He said that the AI model managed to complete a task that they couldn't even understand how it works, making them feel unusable

However, the two GPT-5 initial testers said that the jump from GPT-4 to GPT-5 did not feel as big as the increase from GPT-3 to GPT-4.

However, we don't have to wait too long to prove for ourselves how great the GPT-5 was when it was actually released to the public.