We've scoured all the online leaks about GPT-5, and it's really not easy for Altman and OpenAI to paint a rosy picture this time.

If hype were an Olympic event, Sam Altman would break the medal record.

On X, this is a sharp comment from netizens about the recent aggressive marketing of GPT - 5.

Since OpenAI announced that it had reached the level of an IMO gold medalist, criticism of its high - profile marketing has never stopped.

Almost every day, a little bit of news about GPT - 5 is "casually" released. Sometimes Altman laments on a podcast that he is "useless compared to AI"; sometimes The Information publishes an in - depth report on the numerous internal challenges; and sometimes someone posts a chat screenshot with GPT - 5 on X...

People's appetites have been fully whetted, but the core information about the product remains elusive.

Over the past period, GPT - 5 has seemingly always been the "hottest future product". Since last year, media outlets have been reporting one after another, and rumors about its release date have spread from last year until August this year.

As a result, OpenAI has indeed been refreshing various benchmarks, but the long - awaited GPT - 5 has yet to appear.

Compiled according to OpenAI's official model release announcements

The more OpenAI refreshes the benchmarks, the lower people's excitement becomes. After all, even Musk has to rely on the anime character Ani to gain traffic attention, and how many lists Grok 4 has topped hardly interests anyone.

So, OpenAI seems to have shifted to a new rhythm, maintaining its popularity by constantly "hinting" and "whetting people's appetites".

We have compiled all the recent revelations and reports about GPT - 5, trying to break free from this marketing narrative and see what bets OpenAI is making next after this long - term pre - heating. What are the signals worthy of attention before GPT - 5 truly arrives.

The core upgrade of GPT - 5 unifies the foundation and reasoning models

Different from previous updates, the biggest highlight of GPT - 5 is not a vague improvement in scoring, but rather significant performance improvements in three highly practical fields: multimodality, software engineering, and AI agents.

Realize "complete multimodality" capabilities

This is the biggest and most exciting feature in the latest revelation. According to a transcript of an expert conference call published by blogger @Jukanlosreve on the column SemiconSam, the most prominent feature of GPT - 5 is "a significant improvement in multimodal capabilities, achieving a complete multimodal model".

Screenshot from: https://semiconsam.substack.com/p/script-for-expert-call-nvidia-h20

This means that its ability to process and generate various information flows such as text, images, audio, and video will far exceed that of the past, promising a "Studio Ghibli moment" user experience and is expected to top most evaluation lists after its release.

Significant improvement in software engineering capabilities

GPT - 5 is no longer limited to solving algorithm competition problems. Instead, it can delve into large, old, and complex enterprise - level codebases for modification and maintenance. This is regarded as a powerful weapon for OpenAI to directly challenge Anthropic Claude's dominant position in the programming field.

A tester even said that in their side - by - side comparison, GPT - 5 outperformed Anthropic's Claude Sonnet 4.

Meanwhile, as OpenAI's most important partner, Microsoft employees revealed after testing that GPT - 5 can produce high - quality code without consuming more computing resources.

Image from: https://x.com/SadAlbert10/status/1951306799900139976

Some users on X shared that they tried GPT - 5 for free through the OpenAI API and used it to create an HTML page for a hexagonal vibration test, and the results seemed quite good.

Truly usable AI agents

Stronger reasoning capabilities mean that GPT - 5 can better understand and execute complex multi - step instructions, thus performing multi - step tasks with less human supervision, taking a crucial step towards becoming an "AI agent" capable of autonomously completing tasks.

For example, some media reported that it can follow a set of complex rules to determine whether an automated customer service agent should approve a refund, while previous models required a large number of specific cases to learn this.

The "universal verifier" that conquers the subjective domain

To solve the problem of difficult evaluation of reasoning models, OpenAI has developed a technology called the "universal verifier".

It is worth noting that this work on the "universal verifier" was completed last year by the Super Alignment Team led by Ilya.

In essence, it is another AI specifically responsible for checking and evaluating the quality of the main model's answers, even in scenarios without a standard answer. Therefore, GPT - 5 not only performs well in verifiable fields such as programming and mathematics but also makes progress in more subjective fields such as creative writing and strategic analysis.

This technology is also the key for OpenAI to win a gold medal in the International Mathematical Olympiad.

Image from: https://www.reddit.com/r/OpenAI/comments/1mettre/gpt5_is_already_ostensibly_available_via_api/

Some netizens on Reddit also shared that they obtained access to GPT - 5 through the OpenAI API and used it to generate an image. In the comment section, he mentioned that it took more than 50 seconds to generate this image, and he tends to think that the version he used was not the GPT - 5 mini version.

The images generated by ChatGPT - 4o, on the other hand, are more like stick figures.

To achieve the above - mentioned powerful functions, GPT - 5 relies on its most core change: an intelligent scheduling system that can automatically select the optimal solution for users.

Based on reports from multiple media, different from the previous approach of simply pursuing a larger parameter scale, GPT - 5 aims to integrate the traditional GPT language model with the "o" series of models focused on reasoning (such as o1, o3).

If you, like me, have ever been confused by the dazzling model names such as GPT - 4o, o4, and GPT - 4.1, then the primary goal of GPT - 5 is to end our "choice dilemma".

Models available in ChatGPT, which are good at coding, analysis, handling daily tasks, writing, reasoning, etc.

Some researchers speculate that GPT - 5 may not be a single model but an intelligent "routing system" that can dynamically select the most appropriate model to handle user questions based on their difficulty.

For simple requests: It will call the faster and more cost - effective GPT series of models to achieve instant responses.

For complex problems: Such as STEM (science, technology, engineering, and mathematics) or in - depth programming problems, it will automatically switch to more powerful reasoning models for more in - depth "thinking" to ensure the quality of the answers.

Nick Turley, the head of OpenAI ChatGPT, has this vision: "Our goal is to make ordinary users no longer have to think about which model to use." Instead, the goal is for GPT - 5 to automatically select the best model for users' tasks.

Release, versions, and the final suspense

Time: As early as the end of 2023, some media reported that OpenAI might release GPT - 5 in the spring or summer of 2024. However, in May 2024, OpenAI's official statement at the time was:

Recently, the company has started training its "next - generation frontier model", and it is expected that this new system will reach the "next level of capabilities" on the path to AGI (Artificial General Intelligence).

In 2025, rumors about the release date became more frequent. In June, in an interview video with Altman, he said that GPT - 5 "might be released sometime this summer".

In July, some media reported that in order to comply with the EU's AI regulatory laws, OpenAI would release GPT - 5 at the end of July. Later, The Verge revealed that OpenAI planned to launch GPT - 5 at the beginning of August.

Yesterday, some netizens on X captured a screenshot of Altman saying that he would see everyone in 20 hours, but this tweet was quickly deleted, and it is impossible to confirm whether Altman actually posted it.

With the recent appearance of GPT - 5 on platforms such as Perplexity, Cursor, Microsoft Copilot, Flowith, and the ChatGPT macOS app and website, I think the release time should not be too far off.

Image source: https://x.com/testingcatalog/

More precise information came a few days ago when Wired reported that Anthropic had cut off OpenAI's access to the Claude API.

OpenAI was accused of using the API to test Claude's functions (such as coding and creative writing) and comparing them with its own models, which was regarded as "building a competing product" and a violation of Anthropic's business terms.

OpenAI expressed disappointment, saying that...