OpenAI's GPT-6 Astra, released September 3, 2026, establishes a new benchmark for general-purpose AI with a 1.05-million-token context window, 128K output capacity, and Critical cybersecurity classification. At $10 per million input tokens and $50 per million output tokens, Astra enters a crowded field of frontier models but differentiates itself through extreme context length, verified cybersecurity benchmarks, and built-in computer use capabilities. The launch arrives as enterprises accelerate agent-based workflows and regulators demand verifiable safety metrics.
Context window scaling to 1.05M tokens
The most immediately visible change in GPT-6 Astra is its context window. The model supports 1,050,000 tokens total, composed of 922,000 input tokens and 128,000 output tokens. This represents a substantial expansion over previous generations and enables new classes of applications that were previously impractical or impossible.
A 1.05M context window means Astra can process entire books, multi-file codebases, or lengthy research documents in a single request. The practical implications vary by domain. Legal teams can upload discovery documents, entire case files, and related precedent into a single prompt. Research scientists can feed in experimental data, literature reviews, and analysis notes without truncation. Software engineers can include entire monorepo histories, design documents, and test suites in their context.
OpenAI reports the context window is split into retained conversation space and prompt space, with a premium rate boundary at 272,000 input tokens. Above this threshold, input costs double and output costs increase by 50%. The model also supports tiered rate limits, with Tier 1 allowing 500,000 tokens and 500 requests per minute, and Tier 5 permitting 40,000,000 tokens and 15,000 requests per minute.
The context window is not just a size metric. It enables Astra to maintain long-running conversations without forgetting earlier parts of the discussion. This is critical for multi-turn reasoning tasks, where the model must reference earlier statements, calculations, or decisions. The expanded output capacity of 128,000 tokens means Astra can generate lengthy reports, codebases, or analysis documents in a single pass.
Pricing and tier structure
GPT-6 Astra enters the market at a premium price point. Standard API pricing is $10 per million input tokens and $50 per million output tokens. Cached input costs $1 per million tokens, while cache writes are priced at $12.50 per million. These figures are current as of September 2026 and represent OpenAI's standard tier.
The cached input pricing provides significant savings for repetitive prompts. Developers who frequently include the same context across multiple requests can cache it and pay only $1 per million tokens for subsequent retrievals. This is particularly valuable for applications that process large documents repeatedly, such as document analysis systems, code review tools, or research assistants.
The premium rate boundary at 272,000 input tokens introduces a cost acceleration that developers must account for. Large prompts that exceed this threshold will incur higher per-token costs. OpenAI's tiered rate limits mean that high-volume applications may need to scale their infrastructure to meet the higher throughput requirements of Tier 5.
For enterprises, the pricing structure suggests Astra is positioned as a premium tool for high-value use cases rather than a commodity for routine tasks. Organizations that can justify the cost through improved productivity, accuracy, or capabilities will likely see a positive ROI. Those using Astra for routine or low-value tasks may find the cost prohibitive compared to cheaper alternatives.
Verified cybersecurity benchmarks
GPT-6 Astra comes with a suite of verified benchmarks that position it as a model capable of handling sensitive, high-stakes workloads. OpenAI reports 99.9% accuracy on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 72.6% on OSWorld 2.0, and 100% on ExploitBench. These scores are significant because they cover distinct domains: general reasoning, mathematical problem solving, agent benchmarking, and cybersecurity-specific tasks.
The 99.9% ARC-AGI-3 score places Astra among the top performers on the ARC-AGI benchmark, which measures general intelligence across a range of tasks. This score indicates that Astra can handle novel problems that require reasoning and pattern recognition rather than memorization.
The 97.6% FrontierMath Tier 4 score demonstrates strong performance in mathematical reasoning, a domain that has historically been challenging for language models. This capability is valuable for scientific research, engineering calculations, and financial modeling.
The 72.6% OSWorld 2.0 score reflects performance on agent benchmarking tasks, where the model must interact with an environment to complete objectives. This metric is particularly relevant for autonomous agent workflows, where Astra must navigate systems, execute commands, and achieve goals without constant human supervision.
The 100% ExploitBench score is the most striking cybersecurity metric. ExploitBench measures the model's ability to identify and exploit vulnerabilities in code. A perfect score suggests Astra can assist security researchers, penetration testers, and developers in identifying weaknesses in their systems. This capability is gated behind OpenAI's Daybreak program, which requires a separate application and approval process.
Critical cybersecurity classification
OpenAI classifies GPT-6 Astra as a Critical cybersecurity capability, a designation that signals the model's potential to both assist and be misused in cybersecurity operations. The Critical classification places Astra in the same category as other tools that can generate malware, exploit code, or provide actionable intelligence for attacks.
The Critical designation means Astra's full capabilities are not immediately available to all users. Access is gated behind OpenAI's Daybreak program, which involves an application process and approval workflow. This restriction is intended to prevent misuse while still allowing legitimate security researchers, penetration testers, and developers to access the model's capabilities.
The classification also has implications for enterprise customers. Organizations that use Astra for sensitive workloads may need to navigate additional compliance requirements, reporting obligations, and governance processes. The Critical label may trigger specific security reviews, access controls, and audit requirements depending on the organization's risk tolerance and regulatory environment.
From OpenAI's perspective, the Critical classification provides a framework for managing the model's dual-use nature. The tool can assist defenders by identifying vulnerabilities, writing secure code, and analyzing threat intelligence. It can also assist attackers by generating exploits, crafting phishing campaigns, and bypassing security controls. The Daybreak program and other restrictions are intended to maximize defensive use while minimizing offensive misuse.
Computer use and agentic workflows
GPT-6 Astra includes native computer use capabilities, enabling it to interact with graphical user interfaces, execute commands, and navigate systems. This feature positions Astra as a tool for autonomous agent workflows rather than a passive text generator. The model can open applications, manipulate files, run commands, and complete multi-step tasks without constant human intervention.
The computer use capability is particularly valuable for workflows that span multiple tools and systems. A single agent can orchestrate actions across a development environment, a cloud platform, and internal systems, all while maintaining context and reasoning about the overall objective. This reduces the need for human operators to manually switch between applications and copy-paste information.
Agentic workflows are a growing area of interest for enterprises. Companies are building systems where AI agents handle routine tasks, such as data entry, report generation, or code review, while human operators focus on high-level decision making and oversight. GPT-6 Astra's context window and computer use capabilities make it well-suited for this class of applications.
The model supports file search, code execution, and agent tools out of the box. These integrations mean Astra can work with the developer's existing tooling rather than requiring custom implementations. For example, Astra can search through a codebase, execute tests, and generate pull requests in a single workflow. This end-to-end capability reduces friction and increases the value proposition for teams that adopt the model.
Rollout and availability
GPT-6 Astra is rolling out to ChatGPT Plus, Pro, Business, and Enterprise users. The model is gradually being made available through the API and the ChatGPT interface. Early access may be limited to specific regions or customer segments, with broader availability following over the coming weeks.
The rollout strategy suggests OpenAI is carefully managing the launch to ensure stable performance and gather feedback before widespread adoption, and to address any issues that emerge and adjust pricing or capabilities based on real-world usage patterns.
For enterprises, availability through ChatGPT Business and Enterprise plans means existing customers can access Astra without changing their existing workflows. The model integrates into the ChatGPT interface alongside previous generations, so teams can experiment with Astra and compare it to GPT-5.6 Luna, GPT-5.5, and other models in their existing environment. The API availability is critical for organizations that want to integrate Astra into custom applications and workflows.
Competitive landscape
GPT-6 Astra enters a crowded field of frontier models from Anthropic, Google, Meta, and other providers. Anthropic's Claude Fable 5.1 and Claude Opus 4.8 are widely regarded as leaders in overall intelligence, with Elo ratings that place them at the top of WebDev Arena benchmarks. Google's Gemini 3.8 Flash and 3.8 Flash Cyber offer strong performance at lower price points, making them attractive for cost-sensitive applications.
The 1.05M context window sets Astra apart from many competitors, which typically offer context windows in the 200K to 1M token range. This expansion enables new use cases and positions Astra as a tool for long-running, complex workflows rather than simple text generation tasks.
The Critical cybersecurity classification and verified ExploitBench score differentiate Astra in the security domain. Most other frontier models do not publicly disclose cybersecurity-specific benchmarks, making it difficult for organizations to evaluate their suitability for security-sensitive workloads. Astra's transparent benchmarking provides a clear signal about its capabilities in this domain.
The computer use capability is another differentiator. While other models can be integrated with computer use tools, Astra includes this functionality natively. This reduces the need for custom integrations and provides a more seamless experience for developers and enterprises that want to build agent-based workflows.
Enterprise adoption considerations
Enterprises evaluating GPT-6 Astra should consider several factors beyond raw performance. The pricing structure suggests Astra is positioned as a premium tool for high-value applications rather than a commodity for routine tasks. Organizations must identify use cases where the expanded context window, computer use capabilities, and cybersecurity benchmarks justify the cost.
Security and compliance are critical considerations. The Critical cybersecurity classification means Astra may trigger specific security reviews, access controls, and audit requirements. Organizations must assess whether their existing governance frameworks can accommodate the model's capabilities and restrictions.
Integration with existing workflows is another key factor. Astra must fit into the organization's tooling, data pipelines, and development processes. The model's computer use capabilities and API integrations should align with the team's existing infrastructure rather than requiring a complete rebuild.
The Daybreak program for accessing Astra's full capabilities adds a layer of complexity. Organizations that need Critical-level cybersecurity features must navigate the application and approval process. This can delay access and may not be suitable for teams that need immediate deployment without additional review cycles.
Future implications
GPT-6 Astra's launch signals a shift toward longer-context, more capable frontier models. The 1.05M context window sets a new standard that competitors will likely match or exceed in the coming months. The focus on verified benchmarks and cybersecurity classification reflects increasing regulatory pressure and enterprise demand for verifiable safety metrics, and the emphasis on agentic workflows and computer use indicates that the next generation of AI applications will be more autonomous and integrated into daily operations. As more organizations adopt long-context models, teams will need to rethink how they structure information, how they prompt models, and how they manage the costs associated with large context windows.
Conclusion
GPT-6 Astra represents a significant step forward in general-purpose AI, with an unprecedented context window, verified cybersecurity benchmarks, and native computer use capabilities. The model is positioned as a premium tool for high-value use cases rather than a commodity for routine tasks, and its Critical cybersecurity classification and Daybreak access requirements signal that OpenAI is taking a cautious approach to deploying a model with dual-use potential. For enterprises, Astra offers compelling capabilities for long-running, complex workflows, and those that can identify high-value use cases and navigate the access and compliance requirements will likely see substantial productivity gains.