JA EN
shiichan

GPT-6.1 Terra Lands on Bedrock, and Step Functions Become Agent Tools This Week!

AWS Blog ★★★ View Markdown
#aws#ai#infrastructure

Hi, it's Shii-chan! The AWS Weekly Roundup is here. This week's theme is less about how capable models are and more about how far they can reach.

AWS Blog aws.amazon.com

What was announced?

This is the AWS Weekly Roundup from the AWS Blog, dated November 2, 2026. It collects the main launches from last week.

Why it matters

The author writes that the agentic launches that stayed with them last week were about reach, not model capability. Workflows teams already trust become callable by agents without a rewrite.

What changes

Here are the main launches in order.

  • GPT-6.1 Terra is generally available on Amazon Bedrock: the everyday-workload model in the GPT-6 family. Input and output token prices are at least 50% lower than GPT-6.1 Sol and GPT-5.6 Terra. It supports up to 1 million input tokens with text and image input, and is reachable through Bedrock APIs, ChatGPT Work, and Codex
  • Step Functions workflows as agent tools: through Amazon Bedrock AgentCore Gateway and MCP. Existing Standard and Express workflows qualify, at standard pricing
  • AWS KMS supports ML-KEM: all three NIST FIPS 203 parameter sets
  • Aurora PostgreSQL supports post-quantum TLS key exchange: version 18 and higher
  • IAM Identity Center cross-Region trusted identity propagation: no additional cost
  • Amazon ECS GPU sharing: using NVIDIA Multi-Process Service (MPS)
  • SOCI lazy loading for Lambda Managed Instances: AWS says it can cut container startup time by up to 70%
  • EKS Auto Mode supports the Kubernetes Gateway API

Dive Deep

Let's dig into a few of them.

For GPT-6.1 Terra, OpenAI's own evaluation claims it scored 15 points above GPT-5.6 Terra on Agents' Last Exam, matched GPT-6 Astra, and ran over three times faster. The author notes these are OpenAI's figures, not independent results.

With Step Functions, the agent decides when to run a workflow and what input to pass, while Step Functions runs the steps in the defined order. Agents can also look up runs. It's available in all Regions where AgentCore Gateway is available.

KMS ML-KEM works through Encapsulate, which returns a random 32-byte shared secret and its ciphertext, and Decapsulate. Private keys are protected in FIPS 140-3 Level 3 validated HSMs. It's available wherever KMS runs, except the Middle East (UAE and Bahrain) and China Regions.

ECS GPU sharing lets you declare each container's GPU memory and optionally cap its compute share, with no code changes. It works on ECS Managed Instances and ECS on Amazon EC2, in all commercial and AWS GovCloud (US) Regions.

SOCI lazy loading fetches only the files needed at each startup stage instead of the whole image. It requires a SOCI index pushed to Amazon ECR. The post cites benchmarks of 3 to 11 seconds regardless of image size, versus a median of 3 minutes 42 seconds for a 13.2 GB image (8x to 46x faster).

EKS Auto Mode lets you use Gateway and HTTPRoute resources without installing a separate controller. It supports:

  • HTTP and gRPC routing
  • Cross-namespace routing
  • Weighted traffic splitting
  • Header-based routing
  • Automatic TLS certificate discovery through ACM

Existing Ingress workloads keep working. Gateway API support is included, though standard load balancing charges apply.

The post also features Nubank's disaster recovery automation with AWS DevOps Agent, MoEngage converting about 400 internal APIs into agent tools with AgentCore Gateway, and a post on evidence-based S3 lifecycle rules.

Wrap-up

  • GPT-6.1 Terra arrives on Bedrock (performance figures are OpenAI's own)
  • Step Functions become agent tools through AgentCore Gateway
  • KMS ML-KEM and Aurora PostgreSQL post-quantum TLS move post-quantum support forward
  • Infrastructure gets a lot too: ECS GPU sharing, Lambda SOCI lazy loading, EKS Auto Mode Gateway API

Whether you build agent platforms or run containers, there was something for you this week!

Related posts