GPT-6 (Astra, Sol, Luna, Instant)

What it is

GPT-6 is OpenAI's model family that followed GPT-5.6. Its first model, GPT-6 Astra, was unveiled on 3 September 2026 in a limited preview, with public availability on 4 September 2026.^[1] Later members of the family are GPT-6 Sol, GPT-6 Luna and GPT-6 Instant.^[1][4][5]

Origins

GPT-6 is the successor of GPT-5.6. OpenAI said Astra was the product of “years of research and big bets.”^[2] Rollout began with companies in OpenAI's Daybreak cybersecurity program. ChatGPT Plus, Pro, Business and Enterprise users, the API and Amazon Web Services followed.^[2]

Astra went through a formal review with the Trump administration before release, Sam Altman said.^[1][2]

Aidan Clark said it was the first time OpenAI had pre-trained a model on more than 100,000 GPUs, at its Stargate site in Texas. He called it “by far” the company's largest training run.^[1]

Main features

OpenAI said Astra was state of the art in computer use, software engineering, professional work and science. It was also better at staying on task, respecting task boundaries and running multi-step workflows.^[1][2] Computer use means a model operates a computer interface, such as clicking and typing, to carry out work. A multi-step workflow is a task made of several linked actions.

Astra's API specifications list a context window of 1,050,000 tokens and up to 128,000 output tokens. Its knowledge cutoff is 30 April 2026. It is priced at $10 per million input tokens and $50 per million output tokens.^[3]

GPT-6 Sol and GPT-6 Luna followed on 22 September 2026. Luna is OpenAI's most efficient model and costs $0.10 per million input tokens and $0.50 per million output tokens.^[1][4]

GPT-6 Instant, which has an “Intelligent UI” of interactive widgets, rolled out to paid ChatGPT users on 7 October 2026 and to free users the next day.^[1][5]

GPT-6-Cyber, for investigating and reproducing vulnerabilities in authorised environments, was announced on 5 November 2026.^[1]

Safety and controversy

Astra was the first OpenAI model to reach the company's internal “Critical” cybersecurity threshold. Paid users initially got a restricted version that refused some prompts, such as those about cybersecurity.^[1][2]

Reports citing an anonymous insider said Astra uses a “recurrent depth” or looped-transformer technique. The reports said this obscures some of the model's chain of thought, raising monitorability concerns among safety experts.^[1]

Greg Brockman suggested that Astra could eventually be seen as the arrival of AGI.^[1][2]

Its place in the timeline

The Astra launch came after OpenAI paused some research and training following the July 2026 breach of Hugging Face. Reuters reported the launch came amid growing scrutiny over the safety of AI agents.^[2][3]