Preparing for GPT-5: What we know, what to expect, and what’s rumored

OpenAI is getting ready to release GPT-5. Over the past few months, details have come from a mix of official statements, industry analysis, and frequent leaks. Some of the information is reliable, and some isn’t. But together, they offer a rough outline of what to expect.

This article looks at what OpenAI has confirmed, what experts predict, and what recent leaks suggest about GPT-5.

What OpenAI has said publicly about GPT-5

To get a more realistic sense of what GPT-5 is, it helps to start with what’s been confirmed so far. OpenAI has been quite careful in how it discusses the model, but there’s still quite a bit we do know.

Statements from CEO Sam Altman

OpenAI CEO Sam Altman has confirmed the next model will be called GPT-5 and says it’s coming “soon.”

In a July 2025 podcast, he described giving it a problem he couldn’t solve himself. GPT-5 solved it right away, and Altman said it made him feel “useless relative to the AI.” This lines up with an earlier quote where he described GPT-4 as “the dumbest model any of you will ever have to use again, by a lot,” suggesting GPT-5 will be a major upgrade.

At the same time, Altman has raised concerns about risk. He’s compared AI development to the Manhattan Project and said he’s “scared” and “nervous” about how models like GPT-5 could be misused, for example, to create malware, generate phishing scams, or assist in developing chemical or biological weapons. The point, he says, isn’t that the model will act on its own, but that these risks need to be taken seriously.

Safety and testing plans

OpenAI is building GPT-5 under its public Preparedness Framework, which focuses on testing the model’s safety before release.

A key part of this is red teaming, which means giving the model to outside experts to find weaknesses or ways it could be misused. These tests look at risks like cybersecurity threats or harmful content generation.

The goal is to have security and safety features in place before the model goes public and becomes widely available. This kind of safety testing is now standard for major OpenAI releases and is especially relevant for GPT-5, given how widely it’s expected to be used.

What experts and analysts say GPT-5 will likely do

While OpenAI controls the official story, industry analysts, AI researchers, and journalists have been busy trying to piece things together. Here are the most widely agreed-upon expectations.

Real multimodality, including video and audio

GPT-4o could already understand images and voice. GPT-5 is expected to handle full video and audio files on its own. This means it won’t just read subtitles or look at individual frames, but it may be able to understand what’s happening in an entire video, including sound, movement, and context.

With such an update, users could interact with GPT-5 using a mix of text, images, voice, and video while switching between them naturally in one conversation. It would also connect earlier OpenAI tools: Sora, which creates video from text, and Whisper, which turns speech into text.

Better reasoning & more memory

GPT-5 is expected to be better at reasoning, so it should be able to solve harder problems, especially ones that take more steps. It should also make fewer mistakes, which is something that the previous models have struggled with (for example, so-called hallucinations).

Some early testers (though not officially confirmed) have said GPT-5 handles complex reasoning almost at a PhD level. Whether that’s true or not, the goal is to make the model more accurate and better at following logic.

A big part of that could be a larger context window. That’s how much information the model can remember at once in a single prompt. Experts think GPT-5 might be able to handle up to 1 million tokens, which is enough to read and use an entire book, a long research paper, or a big part of code all at once.

AI agent autonomy

GPT-5 is expected to play a big role in OpenAI’s plans for AI agents, which are tools that can actually do tasks for you. ChatGPT already has an early version of this with its Agent tool, which can browse the internet, run code, and connect with other apps.

With GPT-5, these agents are expected to get much better. Instead of waiting for step-by-step instructions, the model could handle more complex tasks on its own, including things like planning a full trip with multiple stops, or managing a series of tasks across different software for work.

Leaks, rumors, and unconfirmed info

There are a lot of rumors about GPT-5, just like with past OpenAI models. While they haven’t been confirmed, some match with what OpenAI has done before or come from sources that seem hard to fake.

Public tests and API leaks

Throughout 2025, a few mysterious models popped up on public AI benchmark sites, using codenames like “Zenith,” “Summit,” and “Nectarine.” They performed better than any known OpenAI models, and then quickly disappeared. Many suspect these were early versions of GPT-5 being tested quietly in the open.

On Reddit, one developer said they briefly accessed an API endpoint labeled gpt-5-bench-chatcompletions-gpt41-api-ev3. According to their post, the model quite effortlessly handled a complex coding task that GPT-4o couldn’t complete.

And then, toward the end of July, something similar to past model launches happened: parts of a GPT-5 documentation page were briefly indexed by Google before being taken down. Around the same time, a few developers noticed a hidden “gpt-5-preview” model listed inside the API. None of this confirms anything on its own, but together, it looks like OpenAI is getting everything in place for a launch.

Tech specs

One of the most talked-about technical rumors is that GPT-5 uses something called a Mixture-of-Experts (MoE) setup. Instead of one big model handling every request, it’s made up of many smaller “expert” parts (only a few of which run at a time, depending on what the input needs).

This allows for the model to be much larger overall, without slowing it down or making it too expensive to run.

Cost & training data

Some AI researchers say most high-quality public data for training has already been used up. That leaves two likely sources for GPT-5: large private datasets that OpenAI has licensed, and synthetic data produced by earlier models, chosen to help the new model learn in more targeted ways.

Training a model this size is expensive. Some reports estimate that a full training run for GPT-5 could cost over 500 million dollars. OpenAI hasn’t confirmed that number, but if it’s even close, it shows the scale of compute involved.

Microsoft provides the infrastructure behind that, with reports pointing to more than 100,000 NVIDIA H100 GPUs running in its cloud.

When will GPT-5 be released & availability

OpenAI seems to be following a familiar release pattern, and there are enough signs now to make a pretty good guess about timing.

Most people watching this space expect GPT-5 to be released sometime in August 2025. That’s based on a few clues, like infrastructure updates from Microsoft, comments from people close to the project, and even a GPT-5 help page that briefly showed up on Google before being taken down.

Some developers think it’ll happen in the first or second week of August. But OpenAI has delayed launches before when final safety tests weren’t done yet, so that’s still possible. Even so, all signs right now point to an August release.

Phased out release

The model’s launch will almost certainly follow a phased release:

1. Private preview to enterprise partners and key API customers.
2. Public access for paying ChatGPT subscribers (Plus, Pro, Team, Enterprise).
3. Gradual rollout to the broader developer API and, eventually, to some level of free use for general users.

Releasing GPT-5 in stages helps OpenAI catch problems early and keep the system stable as more people start using it. It also ties into Sam Altman’s idea of eventually making GPT-5 free for everyone. If that happens, it would need to roll out gradually because it would cost a lot and be hard to manage all at once.

FAQs

1. What is GPT-5?

GPT-5 is the next major language model from OpenAI. It’s officially confirmed and will be a “unified model,” meaning it combines reasoning and multimodal features (like handling images, audio, and video) into a single system.

2. Is GPT-5 really a big upgrade over GPT-4?

Based on statements from OpenAI leadership and early reports, GPT-5 is expected to be more than a small upgrade. CEO Sam Altman has described GPT-4 as “the dumbest model any of you will ever have to use again,” suggesting GPT-5 will be noticeably more capable, especially in reasoning and handling more complex tasks.

3. When will GPT-5 be released?

Most signs point to a release in August 2025. Some developers expect it could come as early as the first or second week of the month, though the exact timing depends on final safety testing.

4. How will GPT-5 be rolled out?

OpenAI plans a phased release. It will start with private previews for enterprise customers and API partners. The next stage will be access for paying ChatGPT subscribers (Plus, Team, Pro, and Enterprise plans). Wider access will follow later.

5. Will there be different versions of GPT-5?

Yes. GPT-5 is expected to be released in multiple versions, similar to past models. There may be a full-featured model, a smaller “mini” version, and a lightweight “nano” version for faster or more affordable use. This helps OpenAI serve a range of users and use cases.

6. What has OpenAI said about it?

CEO Sam Altman says GPT-5 will be a big step up from earlier models. But he’s also said he’s worried about how it could be misused and believes it’s important to have strong safety measures.

7. How is OpenAI handling safety?

GPT-5 is being developed under OpenAI’s Preparedness Framework. This includes internal testing and external “red teaming,” where outside experts try to find problems or security risks in the system before it launches.

8. What new features can we expect in GPT-5?

Experts expect several upgrades:

• It will likely support video and audio natively, not just text and still images.
• It may have a dedicated reasoning engine to handle more complex tasks with fewer mistakes.
• Its context window could expand to up to 1 million tokens, allowing it to work with much longer documents or code.
• It’s also expected to improve AI agent tools, making them better at completing complex tasks without step-by-step instructions.

9. What are GPT-5 AI agents?

AI agents are features that let the model do more than just answer questions. GPT-5 is expected to support more advanced agents that can browse the web, run code, and work across apps to handle multi-step tasks with less human input.

10. What’s rumored but not confirmed?

GPT-5 may use a Mixture-of-Experts (MoE) architecture, which means it’s made up of many smaller models, and only some are used at a time, which helps with performance and efficiency.

It’s believed that GPT-5 was trained on a mix of private and synthetic data, since high-quality public data has mostly been used up.

Some reports suggest the total cost of training GPT-5 may have exceeded $500 million.
There have also been unconfirmed sightings of GPT-5 in public benchmarks, API endpoints, and documentation leaks, which many believe are early versions being tested.

Want to learn more about large language models, including Google’s Gemini, which has held the top spot on several benchmarks since its launch? At Revolgy, we work with teams to understand how these models work and how they can be used in real situations. If you’re figuring out what tools make sense for your business, we can help, contact us today.