GPT-6 Astra on Amazon Bedrock: enterprise AI agents with hardened security and governance

OpenAI's GPT-6 Astra is now generally available on Amazon Bedrock, giving AI builders a production path for agents that reason across code, documents, and browser workflows without sacrificing AWS-level security and governance. The pairing matters because it combines Astra's deeper judgment with Bedrock's zero-operator access, encryption, IAM controls, and data protections. You can call the model directly through the Bedrock APIs or configure ChatGPT Work and Codex to use it.

Deeper reasoning for the hardest production tasks

GPT-6 Astra is designed for problems where you need to reconcile conflicting inputs, trace dependencies, and decide what to prioritize. The AWS blog post highlights financial analysis (cross-checking contradictory data sources), contract review (scanning hundreds of pages within a context window of up to 1 million input tokens), and codebase investigation (reasoning through dependencies from diagnosis through testing). These aren't just benchmark claims: the model is built for the kind of multi-step work that production agents actually do.

Security and governance on Bedrock

For builders shipping regulated workloads, the security story is the main event. Amazon Bedrock enforces zero-operator access at the chip level, so even AWS operators cannot see prompts or completions. Data is encrypted in transit and at rest. Access is governed by IAM policies, every invocation is logged in CloudTrail, and you can route traffic through VPC endpoints using PrivateLink with organization-level data perimeter policies.

Your inference data is not used for model training, and you do not need to opt into sharing data with OpenAI. Classifier-flagged traffic for abuse detection is retained by AWS for up to 30 days, but you can request zero data retention through your AWS account team.

Integrations for code and enterprise workflows

OpenAI introduced new enterprise plugins for ChatGPT Work that extend Astra's browser-use capabilities across Workday, Navan, and Avalara for tasks in data analytics, operations, and finance. These plugins work through existing user accounts and within administrator permissions, so they do not grant Astra additional access.

For developers, Codex is a software engineering agent that works with local files, repositories, terminals, and IDEs. Configured to use GPT-6 Astra on Bedrock, it applies the model's reasoning and computer-use capabilities across investigation, implementation, and testing. Codex is accessible through the ChatGPT desktop app, CLI, VS Code, JetBrains IDEs, and Xcode. The Agent Toolkit for AWS connects Codex to AWS documentation, APIs, and services with a single terminal command.

Prompt caching and cost optimization

For recurring workflows such as document review, codebase analysis, or agents grounded in company standards, GPT-6 Astra supports both implicit and explicit prompt caching. Explicit caching lets you set cache breakpoints to control which context is cached. Reusing that context in subsequent requests reduces repeated processing, cost, and latency.

Caveats and unknowns

Pricing details were not included in the launch blog post. AWS directs users to the Bedrock documentation for pricing, so builders should check region-specific rates before budgeting. The model's Critical cybersecurity classification under OpenAI's Preparedness Framework is a vendor claim; the real-world impact will depend on how automated safeguards perform in multi-step agent workflows. Similarly, the 47% faster task completion claim from one source is not independently verified and may vary by use case.