OpenAI's GPT-6 Astra Launches on Amazon Bedrock With 1M-Token Context Window

GPT-6 Astra Goes Live on Amazon Bedrock

OpenAI's GPT-6 Astra, described by the company as its latest and most capable model, is now generally available on Amazon Bedrock. Enterprises can call it directly through Bedrock APIs or route ChatGPT Work and Codex to run on Astra via Bedrock's infrastructure.

The headline technical spec: a context window of up to 1 million input tokens, enabling the model to process hundreds of pages of contracts or large codebases in a single pass. Pricing for Bedrock inference has not yet been disclosed; organizations should consult Bedrock's documentation for Region-specific rates and inference profiles.

What's New in Astra

According to OpenAI, GPT-6 Astra applies deeper reasoning to tasks that require reconciling conflicting inputs — financial analysis, contract review, and multi-step codebase debugging are cited as target use cases. The model also advances computer and browser use, letting it operate across software interfaces directly when no API or connector exists, and supports both implicit and explicit prompt caching, with configurable cache breakpoints to cut repeated processing, cost, and latency on recurring workflows.

No standard third-party benchmark scores (MMLU, HumanEval, etc.) were disclosed in this announcement; all capability claims described here are OpenAI's own characterizations.

First Model to Hit OpenAI's Critical Cybersecurity Tier

OpenAI evaluated GPT-6 Astra under its Preparedness Framework and classifies it as the first OpenAI model to reach the Critical classification for cybersecurity capability. At this tier, OpenAI states automated safeguards monitor for misuse in real time and can pause or halt activity that crosses defined boundaries, operating within the Bedrock service boundary alongside AWS's own controls.

Enterprise Access and Governance

GPT-6 Astra is accessible three ways: directly via Bedrock APIs for custom applications, through ChatGPT Work (available on the ChatGPT desktop app for Mac and Windows) for document and spreadsheet generation, and through Codex — accessible via the ChatGPT desktop app, CLI, VS Code, JetBrains IDEs, and Xcode — for software engineering tasks including bug fixes, testing, and pull requests.

Alongside the model launch, OpenAI introduced new ChatGPT Work enterprise plugins extending browser-use capabilities to business intelligence tools, Workday, Navan, and Avalara. OpenAI states these plugins operate within existing user permissions and do not grant Astra additional access beyond what administrators configure.

On the infrastructure side, AWS says Bedrock enforces zero-operator access at the chip level, encrypts data in transit and at rest, logs every invocation in AWS CloudTrail, and supports routing through VPC endpoints via AWS PrivateLink. AWS states inference data is not used for model training and customers are not required to opt into data sharing with OpenAI. Classifier-flagged traffic for abuse detection is retained up to 30 days; zero data retention can be requested through AWS account teams.

What This Means

This release extends OpenAI's multi-cloud distribution strategy, putting its most capable model behind AWS's compliance and governance tooling — a draw for regulated enterprises that need audit trails and IAM-based access controls rather than direct OpenAI API access. The 1M-token context window and native computer-use capability position Astra for long-document and agentic workflows, but the more notable data point is the Critical cybersecurity classification: it signals OpenAI's own safety framework now rates this model's offensive capability high enough to require real-time monitored guardrails, a first for the company. Absent published benchmark scores or pricing, buyers evaluating Astra against competing frontier models on Bedrock will need to run their own comparisons before committing production workloads.