OpenAI’s GPT-6 Astra Reaches General Availability on Amazon Bedrock

AWS said on September 8, 2026, that OpenAI’s GPT-6 Astra is now generally available on Amazon Bedrock, describing it in an AWS Machine Learning Blog post as OpenAI’s latest and most capable model. Customers can call the model directly through the Amazon Bedrock APIs or configure ChatGPT Work and Codex to use it, and OpenAI is introducing new ChatGPT Work enterprise plugins as part of the launch.

The model runs on the Amazon Bedrock inference engine, which AWS said delivers the scalability and reliability needed for production use, with established AWS controls for securing workloads, governing access, and auditing model invocation activity.

Reasoning, Context, and Caching on Bedrock

AWS said GPT-6 Astra is built for work that involves weighing competing inputs against one another, tracing dependencies, and setting priorities. In financial analysis, the model can spot inconsistencies between conflicting data sources that could alter a recommendation, according to AWS. In contract review, the company said, a context window of up to 1 million input tokens lets the model take in hundreds of pages and flag the provisions that pose the most risk. For software work, AWS said the model can trace a problem across a large codebase, work through its dependencies, and take a fix from initial diagnosis to testing.

AWS also said GPT-6 Astra advances computer and browser use, letting it operate across applications and drive software interfaces directly when no API or connector is available.

For workflows that reuse the same context across requests, such as recurring document review, codebase analysis, or agents grounded in company standards, GPT-6 Astra supports both implicit and explicit prompt caching on Bedrock. Explicit caching lets developers place cache breakpoints that determine which context gets cached, and AWS said reusing that cached context on later requests reduces repeated processing, cost, and latency.

Security, Governance, and Data Handling

OpenAI evaluated GPT-6 Astra through its Preparedness Framework, which assesses model capabilities across safety-relevant domains and applies progressively stronger safeguards as capabilities advance. AWS said Astra is the first OpenAI model to reach the framework’s Critical classification for cybersecurity capability, a level at which automated safeguards monitor misuse in real time and can pause or stop activity that exceeds defined boundaries. Those safeguards operate within the Amazon Bedrock service boundary and complement AWS security controls when the model works across code, software, and tools, according to the post.

AWS said Bedrock enforces zero-operator access at the chip level, meaning AWS’s own operators cannot read prompts or completions during inference. Encryption covers data in transit and at rest, AWS Identity and Access Management policies govern access, and AWS CloudTrail logs every model invocation. Traffic can run through AWS PrivateLink virtual private cloud endpoints, and customers can set data perimeter policies at the organization level to guard against data exfiltration across account and network boundaries.

AWS said customer inference data is not used to train models, and running GPT-6 Astra does not require opting in to data sharing with OpenAI. Traffic flagged by classifiers for automated abuse detection is held by AWS for up to 30 days and processed programmatically, and customers can ask their AWS account team for zero data retention.

ChatGPT Work, Codex, and Enterprise Plugins

AWS described ChatGPT Work as a productivity agent that converts complex business tasks into finished deliverables. Running on GPT-6 Astra, it can pull information from applications and files, browse the web, and produce spreadsheets, slides, documents, and sites. Users decide which applications and websites the agent may reach, manage file uploads and downloads, and can require confirmation before designated actions; they can also watch its progress, redirect it, and approve key steps along the way.

ChatGPT Work is available through the ChatGPT desktop app for Mac and Windows. AWS said the new enterprise plugins carry Astra’s browser-use capabilities into business intelligence tools, Workday, Navan, and Avalara, covering data analytics, operations, and finance work, and that the plugins run through existing user accounts within administrator-set permissions rather than giving Astra added access.

Codex, the software engineering agent, works across local files, repositories, terminals, developer tools, and development environments, writing features, fixing bugs, running tests, and opening pull requests. AWS said that when Codex is configured to use GPT-6 Astra on Amazon Bedrock, it puts Astra’s reasoning and computer-use capabilities to work across investigation, implementation, and testing. Codex is accessible through the ChatGPT desktop app, a command-line interface, VS Code, JetBrains IDEs, and Xcode, and the Agent Toolkit for AWS gives Codex access to AWS documentation, APIs, and service capabilities through a single terminal command.

OpenAI’s Reported Benchmarks, Pricing, and Safeguards

In its own announcement, OpenAI introduced GPT-6 Astra as what it called the world’s most intelligent and aligned model, reporting state-of-the-art results in computer use, browsing, software engineering, cybersecurity, science, and professional work. Its reported scores include 99.9% on ARC-AGI-3, 100% on ExploitBench, a 98% score on FrontierMath Tier 4, and 57.9% on Terminal-Bench 4.0, a result OpenAI presented against 37.3% for GPT-5.6 Sol. Greg Kamradt of the ARC Prize Foundation said in comments published with OpenAI’s announcement: “On ARC-AGI-3, Astra surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark.”

OpenAI said Astra will decline more advanced cybersecurity work, such as producing proof-of-concept exploits for vulnerabilities. The company said it plans to broaden access through its Daybreak program in the coming weeks with less restrictive safeguards, opening defensive workflows that include vulnerability and proof-of-concept validation, malware analysis, and detection engineering.

OpenAI also said it is deploying misalignment monitoring in production for Astra-class models, using classifiers that inspect the model’s reasoning and actions for unauthorized behavior and halt activity that appears unauthorized. OpenAI said its evaluations found Astra’s written reasoning harder to monitor than GPT-5.6 Sol’s, a finding it attributed to the model’s tighter control over written reasoning on simpler tasks and its ability to solve problems in fewer written steps, and said improving monitorability remains a research priority.

OpenAI lists Standard API pricing for the model, available as gpt-6-astra, at $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes and a Fast mode it said runs up to 2x the speed of Standard processing at 2x the Standard price. OpenAI said Astra is rolling out first to a limited set of organizations, with availability for all ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, and developer access through the OpenAI API, Microsoft Azure, and Amazon Bedrock.

AWS’s OpenAI on Amazon Bedrock page lists GPT-6 Astra among the OpenAI frontier models on the service, a lineup that also includes GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, and GPT-5.4, and presents Astra as the largest and most capable OpenAI model to date. Customers can get started in the Amazon Bedrock console or programmatically through supported Amazon Bedrock APIs; AWS documentation covers supported regions, endpoints, APIs, features, inference profiles, and pricing.