What Is GPT-6, and How Do Large Language Models Actually Work?

GPT-6 is not a single product launched on one day. OpenAI released it in two stages: a frontier-level version called GPT-6 Astra in early September 2026, followed by two consumer-facing versions, GPT-6 Sol and GPT-6 Luna, that replaced the previous model inside ChatGPT in October 2026.

Astra first, then Sol and Luna for everyone

According to OpenAI's own account summarised on Wikipedia, GPT-6 Astra went into limited preview on September 3, 2026, and reached paying ChatGPT users the next day. OpenAI described it as its most capable system to date, pointing to gains in coding, mathematics and autonomous computer use. OpenAI vice president Aidan Clark said the model was the company's largest training effort yet, calling it the first time OpenAI had pretrained a model on more than 100,000 GPUs, at its Stargate site in Texas.

OpenAI classified Astra at the 'Critical' cybersecurity capability level under its internal Preparedness Framework, according to DataCamp's review of OpenAI's published materials. DataCamp reported that this rating reflects Astra's capacity to help create cyber exploits rather than an ability to autonomously discover unknown flaws in hardened systems on its own, and that OpenAI restricted the model's more advanced cybersecurity tasks, such as producing proof-of-concept exploits, behind its Daybreak access program at launch. OpenAI president Greg Brockman told reporters he thought it was reasonable to say the industry had entered what he called an AGI era, though he stopped short of saying OpenAI had achieved artificial general intelligence.

On October 7, 2026, OpenAI brought a second wave of GPT-6 into the main ChatGPT app, this time built for everyday conversation rather than specialised agentic work, according to OpenAI's announcement on its developer community forum and reporting by TechCrunch. Plus, Pro, Business and Enterprise subscribers received a version called GPT-6 Sol immediately; free and Go-tier users began getting GPT-6 Luna the following day.

The visible change: an 'Intelligent UI'

Alongside the model swap, OpenAI introduced a feature it calls Intelligent UI. Instead of answering only in paragraphs, ChatGPT can now assemble a reply from text, charts, tappable buttons, forms and small interactive tools such as calculators, according to OpenAI's community post and TechCrunch's reporting. OpenAI product manager Aarush Selvan said the goal was to make complex topics easier to understand, explaining that the most helpful answers to many questions are not plain text alone, as reported by TechCrunch. The system decides case by case whether a diagram, an interactive chart or plain text best fits a question, and users can adjust how often these visual elements appear.

What OpenAI's own safety testing found

OpenAI's system card for the October release, published on its Deployment Safety Hub, says GPT-6 Sol and Luna improved on several safety measures compared with the prior GPT-5.6 models: stronger resistance to jailbreak attempts, including multi-turn attacks, less dishonest or deceptive behaviour, and better factual accuracy on most metrics. The same document also reports statistically significant regressions in how the models handled a narrow set of sensitive evaluation categories, including self-harm-related prompts, with Luna showing additional regressions on gore and sexual content in internal testing; OpenAI's own assessment described the problematic outputs it reviewed as low-severity and still within its general safety expectations.

Under OpenAI's Preparedness Framework, the system card also classifies both Sol and Luna as 'High' capability in biological/chemical and cybersecurity risk categories, but below the 'Critical' threshold — a lower risk tier than the Astra model released the previous month, which OpenAI had rated Critical for cybersecurity.

How large language models actually work

GPT-6, like GPT-4, Claude or Llama, is built on a neural network architecture called the transformer, introduced by Google researchers in 2017, according to IBM's research explainer. A transformer breaks text into small units called tokens, converts each token into a numerical vector, and then uses a mechanism called self-attention to work out how strongly every token relates to every other token in the input — including ones far apart in the text.

Before any of that happens, the model has to be trained. IBM's explainer describes this as starting with massive datasets — billions or trillions of words drawn from books, websites and code — which are cleaned, tokenised and fed through the network using self-supervised learning, a method that does not require humans to label the correct answer for every example.

Once trained, the model works by taking a prompt, converting it to tokens, and repeatedly predicting the single most likely next token, one at a time, until the response is complete. As IBM's explainer puts it, the system does not know its final answer in advance; it generates the response piece by piece using the statistical relationships learned during training, rather than retrieving a stored fact or reasoning the way a person does.

What this means for trusting the output

Because the system is predicting plausible text rather than verifying truth, IBM's explainer lists hallucination — confidently generated but false information — as an inherent limitation of the architecture, alongside the risk that biases present in training data get reflected or amplified in outputs. Training and running models at this scale also demands very large amounts of computing power and energy, a point OpenAI itself underscored when it said Astra's training run used over 100,000 GPUs at a single site.