OpenAI's GPT-6 Astra becomes generally available on Amazon Bedrock

OpenAI’s GPT-6 Astra is now generally available on Amazon Bedrock, Amazon’s managed service for running AI models through one interface. Astra is a reasoning model, meaning it works through problems in steps before it answers. AWS describes it as OpenAI’s latest and most capable model, and says customers can call it directly through the Amazon Bedrock APIs.

There are other ways to run it. Teams can configure ChatGPT Work, OpenAI’s agent for turning business tasks into finished documents, or Codex, its software engineering agent, to use GPT-6 Astra on Bedrock. Alongside the launch, OpenAI is adding enterprise plugins for ChatGPT Work that extend the model’s browser use across business intelligence tools, Workday, Navan, and Avalara. AWS says the plugins work through a user’s existing account and within permissions set by administrators, so they do not grant the model extra access.

AWS frames the model as suited to work that involves weighing competing inputs. For financial analysis, it says Astra can reconcile conflicting data sources and identify discrepancies that could change a recommendation. For contract review, it can read hundreds of pages within a context window of up to 1 million input tokens (the text a model can weigh at once) and point to the provisions that carry the most risk. Across a codebase, AWS says the model can trace dependencies and carry a fix from diagnosis through testing, and can operate software directly when no API or connector is available.

For workflows that reuse the same context across requests, GPT-6 Astra supports prompt caching, which reuses stored parts of earlier prompts. With explicit caching, a developer sets breakpoints to control which context is held. AWS says reusing that context cuts repeated processing, cost, and latency.

On safety, AWS says GPT-6 Astra is the first OpenAI model to reach the “Critical” classification for cybersecurity capability under OpenAI’s Preparedness Framework, the company’s system for assessing model risk. At that level, AWS says, automated safeguards watch for misuse in real time and can pause or stop activity that crosses defined limits. It says its own controls run alongside these: prompts and completions cannot be reached by AWS operators during inference, data is encrypted in transit and at rest, access is governed by AWS identity policies, and every model call is logged in AWS CloudTrail. Inference data is not used to train models, and using Astra does not require opting into data sharing with OpenAI. Traffic flagged by abuse classifiers is kept by AWS for up to 30 days, and customers can request zero retention through their account team.

The announcement does not list prices or the AWS Regions where the model runs, pointing instead to the Bedrock documentation for pricing, Regions, endpoints, and inference profiles. ChatGPT Work runs through the ChatGPT desktop app for Mac and Windows, and Codex is available through that app, a command-line tool, VS Code, JetBrains IDEs, and Xcode. Any expansion of the ChatGPT Work plugins beyond the four named services would appear in that documentation and in OpenAI’s release notes.