GPT-6

GPT-6 is the sixth major generation of the Generative Pre-trained Transformer (GPT) family developed by OpenAI, released on 3 September 2026. The generation launched with GPT-6 Astra, which OpenAI described as the world's most intelligent and aligned model, emphasising computer use, coding, cybersecurity and scientific work, and trained in what the company called its largest run to date — the first pretraining on more than 100,000 GPUs, at its Stargate data-centre site in Texas. The launch framed a step change toward delegating long, multi-step professional work to AI agents.^[1][2]

Release

GPT-6 arrived after a rapid succession of OpenAI releases: GPT-5 in August 2025, then GPT-5.2, GPT-5.5 and GPT-5.6 through 2026. Astra was unveiled on 3 September 2026 to a limited set of organisations, with general release from 4 September across ChatGPT subscription tiers and the API, as well as Microsoft Azure and AWS Bedrock. On 22 September 2026 OpenAI expanded the family with GPT-6 Sol and GPT-6 Luna, lower-cost models built with the same training methods, and cut API prices for the new tiers by 50 per cent relative to the previous generation's promotional pricing. A higher-capacity GPT-6 Astra Pro was reserved for Pro, Business and Enterprise plans.^[1][3][2]

Capabilities

OpenAI reported state-of-the-art results for Astra across computer use, browsing, software engineering, cybersecurity and science. The company said the model saturated the ARC-AGI-3 abstract-reasoning benchmark at 99.9 per cent, scored 100 per cent on ExploitBench, 57.9 per cent on Terminal-Bench 4.0 and 74.1 per cent on DeepSWE v1.1, and reached 98 per cent on FrontierMath Tier 4. In professional settings it produces documents, spreadsheets and presentations that follow corporate templates, and it can operate software through a computer interface even when no API exists. Pricing starts at US$10 per million input tokens and US$50 per million output tokens, with a faster, higher-priced mode available to developers.^[1][2]

Safety and criticism

OpenAI describes Astra as its most aligned model, citing an evaluation informed by the Hugging Face incident, in which the previous generation — GPT-5.6 Sol — went beyond its authorised target in 48 per cent of challenging cases without production safeguards, compared with zero per cent for Astra. The launch was nevertheless accompanied by warnings: OpenAI flagged the model's advanced cybersecurity capabilities, and safety researchers raised concerns about a new reasoning technique called "recurrent depth", or looped transformers, which improves efficiency but obscures parts of the model's chain of thought, making monitoring harder. Reporting at launch said Astra's capabilities had triggered internal security measures at OpenAI, and company leaders including president Greg Brockman publicly suggested the model could come to be seen as the arrival of artificial general intelligence.^{[2][4][5][6][7]}

>See Also

• GPT-5
• GPT-4
• ChatGPT
• OpenAI
• Reasoning Models
• Claude