AWS This Week: CloudWatch Omni, GPT-6 Sol & Luna and Claude Opus 5.5 on Bedrock (Video)

The week’s AWS announcements that actually matter for architects — in 88 seconds, with the “so what” for your architecture. This week the theme is the AI platform: a new observability service built for agents, and two frontier model families arriving on Bedrock the same day. Watch first, details below.

This week’s picks

1. Amazon CloudWatch Omni is generally available — an evolution of CloudWatch organised around your teams and the applications they run, rather than around resources. You create spaces in a central account that see telemetry across your AWS accounts and Regions — and other clouds, Azure workloads included. Omni discovers services, maps dependencies and surfaces golden metrics automatically; it is OpenTelemetry-native, and you drive it through chat (root cause powered by AWS DevOps Agent), through a guided point-and-click path, or from your own tooling via Agent Toolkit for AWS. It is delivered off-console: a standalone web experience behind your SSO, plus an IDE extension. Architect’s notes: (a) GA in US East (N. Virginia), US West (Oregon) and Europe (Ireland) only — the choice of the central account and Region is a data-residency decision, take it deliberately; (b) pricing is standard CloudWatch rates for ingestion and retention, with the Omni analysis layer charged per GB scanned ($0.005/GB in N. Virginia) and query volume up to five times your monthly ingestion included free — dashboards and alerts included, first centralised copy free, each additional copy $0.05/GB. (announcement, pricing, News Blog on the application side)

2. Omni for agents: evaluations move into the trace — the same launch carries a dedicated agent-observability experience with an evaluation-driven development workflow. Every prompt, model call and tool invocation is captured as a trace; built-in evaluators score correctness, coherence, retrieval quality and tool selection; you compare prompt versions side by side, build test datasets from production traffic, run experiments across configurations and detect regressions automatically — across LangGraph, CrewAI, OpenAI Agents SDK, Vercel AI SDK and Strands. Developers get a free extension for VS Code, Cursor and Kiro that instruments, debugs and evaluates agents locally, with no AWS account required; operators get the standalone web experience — and it is the same trace on both sides. The structural point: the eval set stops being a notebook someone runs before a release and becomes a first-class object of the observability stack, which is exactly what the industry signal below needs. (announcement, News Blog on the agentic side)

3. OpenAI GPT-6 Sol and GPT-6 Luna are generally available on Amazon Bedrock — Sol and Luna join Astra, so the GPT-6 family on Bedrock now has three tiers. Sol is the daily model for recurring complex tasks and software development (on OpenAI’s internal factuality evaluation it makes roughly half as many mistakes as GPT-5.6 Sol); Luna is the efficient model for high-volume summarisation, extraction, classification and routing, with adjustable reasoning effort to trade quality against speed and cost per request. Both support up to 1M tokens of context, and both sit behind the Bedrock controls you already run: IAM, invocation logging and audit, Guardrails. Architect’s note: the routing pattern — Luna for the bulk, Sol for the hard cases, Astra for the rare ones — is now available inside one API surface, so the model tier becomes a per-request decision rather than a platform choice. (announcement, Bedrock OpenAI model documentation)

4. Claude Opus 5.5 is available on AWS — the first model of the Claude 5.5 family. According to Anthropic it completes tasks using fewer tokens than Claude Opus 5, at a lower price per token, with cheaper cache reads stacking on top of the efficiency gain, and it thinks adaptively, deciding how much effort each request needs. Two doors on AWS: Amazon Bedrock (zero data retention by default, regional data residency, Guardrails and Knowledge Bases) or Claude Platform on AWS. Architect’s note: “fewer tokens per task” is the number to verify on your workload — token efficiency and unit price move independently, and a cheaper model that talks more can cost more. (announcement, Anthropic’s launch post)

Industry signal — frontier model prices fell twice in one day. Anthropic priced Opus 5.5 at $4 per million input tokens and $20 per million output tokens, saying it costs 40% less to run than Opus 5 on a typical workload while performing at the level of Claude Fable 5.1 on most work; the same day, OpenAI halved GPT-6 Sol against GPT-5.6 Sol ($4 → $2 input, $20 → $10 output) and cut Luna by half or more ($0.20 → $0.10 input, $1.20 → $0.50 output) — the pricing I analysed in August is already the old baseline. Your model bill is a moving target, and the temptation to swap models on a price signal is exactly where quality regressions hide. Treat every model swap as a release: an eval set that defines “better” for your use case and a rollback criterion, both written down before the swap — item 2 above is the tooling for it. (Anthropic, OpenAI, coverage)

Which one changes your roadmap? The video also runs on LinkedIn — previous edition: September 7. If Omni caught your eye, I wired AWS DevOps Agent into an observability-to-ITSM loop earlier this summer; and if model swaps run through a gateway in your platform, read which of the two gateway jobs it actually does.