GPT-6 Astra lands on Amazon Bedrock: 1M context, in your AWS account
OpenAI's flagship now runs inside Bedrock — IAM, VPC and audit trails around it, chip-level isolation, no training on your prompts; Codex can point at it too.
Copy markdown
Astra, now inside your AWS account
As of Sept 8 you can call GPT-6 Astra straight from the Bedrock console and APIs — same IAM, VPC, and audit trails as any Bedrock model, no separate OpenAI contract. Codex and ChatGPT Work can be pointed at the Bedrock-hosted model, so your terminal and editor stay in one billing and governance boundary.
1M-token context, and it caches
Astra takes up to 1,000,000 input tokens — a whole codebase, a few hundred pages of contracts, or a long agent run in a single call. Implicit and explicit prompt caching cut the latency and cost of re-sending that context across turns.
Computer use and agents, with guardrails
The model ships browser/computer use and multistep agentic workflows; on Bedrock you wrap it in the Agent Toolkit for AWS and enforce access governance and audit trails around it. That's the piece that makes autonomous agents shippable in a regulated stack, not just a demo.
Your prompts stay yours
Zero-operator access is enforced at the chip — even AWS operators can't read your prompts or completions during inference. Data is encrypted in transit and at rest, isn't used for training, and zero data retention is available on request; abuse-detection logs default to 30 days.
Elsewhere: a free agentic model on OpenRouter
Nex-AGI put Nex-N2.5-Pro and -Mini up free on OpenRouter with a 262K context, tuned for verify-and-retry coding and computer-use agents. Self-reported scores are strong (BrowseComp 89.7, SWE-Bench Pro 61.2, Terminal-Bench 82.7); the Apache-2.0 weights are marked “coming soon,” so try it via the API today and treat the benchmarks as the lab's own.