GPT-5 and GPT-6 Astra are successive releases from OpenAI, marking significant advances in Large language model technology and multimodal interaction. GPT-5, launched on August 7, 2025, is the fifth generation in the generative pre-trained transformer series, succeeding GPT-4. GPT-6 Astra, which followed, builds on this foundation with a focus on real-time, agentic, and multimodal experiences, aimed at integrating more deeply into daily workflows and devices.
These launches are part of OpenAI's broader trajectory toward artificial general intelligence, as described by CEO Sam Altman during the GPT-5 unveiling. The models are publicly accessible through ChatGPT and Microsoft Copilot, as well as via the OpenAI API for developers. This article details the background, capabilities, architecture, training, use, and limitations of these systems, drawing on public reporting and technical documentation available as of their release dates.
Background
On April 14, 2023, Sam Altman spoke at the Massachusetts Institute of Technology and stated that OpenAI was not training GPT-5 at that time, emphasizing a focus on GPT-4 development. However, on July 18 of that year, OpenAI filed for a "GPT-5" trademark in the United States, and by November 13, Altman confirmed to the Financial Times that work on GPT-5 was underway.
According to The Information, much of the second half of 2024 involved an internal model known as Orion, intended to become GPT-5. That effort reportedly failed to produce a better model, leading to its release as GPT-4.5 in February 2025. By late July 2025, expectations for a GPT-5 release in early August had grown, with The Verge reporting on July 30 that Microsoft was preparing its Copilot chatbot for the new model.
On August 5, 2025, OpenAI released GPT-OSS, a set of two open-weight models with reasoning capabilities, as a precursor. GPT-5 was unveiled during a livestream event on August 7. The subsequent GPT-6 Astra was introduced later in 2025, with less public lead-up but a stronger emphasis on multimodal and agentic interaction, drawing on lessons from earlier deployments.
Capabilities
At its release, GPT-5 achieved state-of-the-art performance on benchmarks testing mathematics, programming, finance, and multimodal understanding. OpenAI reported improvements over predecessors, including faster response times, better coding and writing skills, more accurate health-related answers, and reduced hallucination rates. The model also introduced a "safe completions" approach, aiming to give safe, high-level responses to potentially harmful queries rather than outright declining them, which reduced rejections for harmless information.
GPT-5 was trained to be more critical and less effusively agreeable than prior models. Early testers, days before launch, were impressed by its coding and mathematical problem-solving, noting marked improvement from GPT-4 but not as large a leap as from GPT-3 to GPT-4. Altman called GPT-5 "a significant step along the path to AGI," offering "PhD-level" abilities across a wide range of tasks.
GPT-6 Astra extended these capabilities with enhanced real-time multimodal processing, allowing for more natural voice, vision, and text integration. It supports agentic functions, such as autonomous web browsing and desktop management, building on the architecture introduced in GPT-5. Energy consumption for GPT-5 was not disclosed, but researchers at the University of Rhode Island estimated a medium-length response consumes slightly over 18 watt-hours, akin to an incandescent bulb running for 18 minutes.
Architecture
GPT-5 is a system comprising a fast, high-throughput model, a deeper reasoning model, and a real-time router that selects the appropriate model based on conversation type, complexity, tool needs, and explicit user intent. This unified design addresses Altman's earlier criticism of manual model pickers. The system card defines two fast models (gpt-5-main and gpt-5-main-mini) and two thinking models (gpt-5-thinking and gpt-5-thinking-mini).
Developers accessing the OpenAI API can use the thinking model, its mini version, and gpt-5-thinking-nano, a smaller and faster variant. Adjustable reasoning effort (low, medium, high, or minimal) and verbosity settings (low, medium, or high) are available. ChatGPT users get access to gpt-5-thinking with parallel test-time compute, called gpt-5-thinking-pro.
GPT-6 Astra adopts a similar routing architecture but emphasizes a tighter coupling between perception and reasoning, using Cross-Attention mechanisms to fuse audio, visual, and textual inputs in real time. Its design supports continuous learning from user interactions, although details on its internal model variants remain sparse as of this writing.
Training
GPT-5 is natively multimodal, trained from scratch on text and images simultaneously, unlike GPT-4 which relied on separately trained components. Its training involved three stages: unsupervised pretraining, supervised fine-tuning, and reinforcement learning from human feedback, using a large-scale multilingual dataset of books, articles, web pages, academic papers, and licensed sources.
GPT-6 Astra followed a similar three-stage process but expanded the pretraining corpus to include more diverse audio and video data. OpenAI used reinforcement learning from human feedback to align responses with user intent rett, with additional fine-tuning for safety and reduced hallucination. The company also employed Curriculum Learning to progressively increase task complexity during training, improving generalization across modalities.
As of late 2025, OpenAI has not disclosed the exact computational resources used for training, but industry estimates suggest large clusters of specialized hardware, possibly involving collaborations with AMD or nvidia? suppliers, though no official confirmation exists.
Use
GPT-5 is free for all ChatGPT users, with Plus subscribers receiving higher usage limits and Pro users getting unlimited access, including limited GPT-5 Pro access. With GPT-5, ChatGPT's "Advanced Voice Mode" was replaced by "ChatGPT Voice", which enables more natural-sounding conversations. Standard Voice Mode retired on September 9, 2025. On November 24, 2025, shopping research was added to ChatGPT, claimed to be a mini model post-trained on gpt-5-thinking-mini.
GPT-5 is also integrated into Microsoft Copilot, with Microsoft planning broader product integration. Apple Inc. is reportedly working to incorporate GPT-5 into Apple Intelligence for iOS 26, iPadOS 26, and macOS Tahoe. GPT-6 Astra is available in ChatGPT and via the API, with a focus on mobile and embedded applications, and is expected to power features in devices from Samsung and other partners.
Safety and Limitations
Security research companies, including Neuraltrust and SPLX, claimed to have compromised GPT-5 within its first day of testing, enabling it to generate detailed instructions for manufacturing explosives. These assessments suggest significant security gaps, potentially rendering GPT-5 unsafe for corporate environments without additional safeguards. OpenAI has acknowledged limitations and continues to iterate on safety measures.
GPT-6 Astra faces similar challenges, with early tests indicating improved refusal rates for harmful queries but occasional failures in edge cases. The models also exhibit biases inherited from training data, and OpenAI encourages ongoing external auditing and red-teaming.
Reception and Impact
The launches have sparked debate about the pace of AI development, with some experts calling for more robust regulation. Academic groups like Stanford AI Lab and Berkeley AI Research are studying the models' reasoning abilities; early results suggest progress but not yet human-like comprehension. The releases have also intensified competition with Anthropic and Google DeepMind, which are developing similar systems.
Industry partnerships are expanding; Oracle Cloud and Microsoft Azure provide cloud infrastructure for deployment, while ARM and Qualcomm are seen as potential partners for on-device inference. As of late 2025, GPT-5 and GPT-6 Astra remain central to OpenAI's strategy, with ongoing updates expected.