One Announcement, Three Very Different Meanings of "Retire"
On June 30, 2026, AWS published a quiet what's-new post trimming its service catalog. The headline number is real: roughly a dozen services stop accepting new customers on July 30, 2026, with more on sunset timelines behind them. But "AWS retires a dozen services" is doing too much work in one sentence, because the announcement bundles three completely different situations under one word, and confusing them is how you either panic over nothing or miss the one deadline that matters.
There is also a pattern in the list that most coverage skipped past. Strip out the plumbing and what is left is largely AWS's own first-generation AI and search products: Amazon Kendra, Amazon Q Business, Amazon Bedrock Agents. This is not AWS retreating from AI. It is the opposite, and that distinction is the actual story.
The Three Buckets
Everything in the announcement falls into one of three categories, and your response depends entirely on which one a service you use is in.
| Bucket | What it means | Do you have to act? |
|---|---|---|
| Maintenance mode | Closed to new customers from July 30, 2026. Existing customers keep running, fully supported, with security patches and bug fixes. No new features, ever. | No emergency. Plan off it on your own timeline. |
| Sunset | A real end-of-support date is set. The service will stop being supported and you must migrate before then. | Yes, on a deadline. |
| Already end of support | Ended June 30, 2026. If you are still on it, you are already unsupported. | Yes, now. |
Bucket 1: Moving to maintenance (closed to new customers July 30, 2026)
Amazon Kendra, Amazon Q Business, Amazon Bedrock Agents (renamed Bedrock Agents Classic), Amazon Cognito Sync, AWS Directory Service Simple AD, AWS IoT Device Defender Detect, AWS Mainframe Modernization Self-Managed, AWS Management Console myApplications, AWS Resource Groups Group Lifecycle Events, AWS Service Catalog Application Registry, AWS Systems Manager Application Manager, and a long list of Amazon SageMaker AI features (A2I, Clarify, Debugger, GeoSpatial, Ground Truth, Mechanical Turk, Model Monitor, Role Manager, Studio Lab).
One exception to flag inside this bucket: AWS IoT Device Defender Detect carries its own August 31, 2026 date rather than the general July 30 one. If you use it, check the specifics for that service directly.
Bucket 2: Sunsetting (end-of-support dates announced)
Amazon WorkSpaces PCoIP, Amazon WorkSpaces Pool, AWS Managed Services (AMS) Advanced, AWS re:Post Private, and Amazon SageMaker Profiler. These have real migration deadlines. The WorkSpaces PCoIP one is the item most enterprises will actually feel, covered below.
Bucket 3: Already ended (as of June 30, 2026)
Amazon Chime SDK Carrier Voice Focus and Amazon SageMaker Ground Truth Plus. If you are still running these, you are on unsupported software today.
What "Maintenance Mode" Actually Means
This is the word that causes the most wasted worry. Maintenance mode is not a shutdown and not a countdown. Using Kendra as the concrete example, because AWS documented it clearly: Kendra entered maintenance mode on June 30, 2026, stopped accepting new customers on July 30, and for existing customers it "remains fully supported" with continued bug fixes and security updates. What stops is new feature development.
So a maintenance-mode service does not break on a date. What you lose is a future. The product is frozen, no new capabilities are coming, and you cannot start a new project on it. That is a strategic signal, not an operational fire. The correct response is to stop treating it as a foundation for anything new, and to plan a migration on a timeline you control rather than one AWS imposes.
The trap is the opposite of urgency: because nothing breaks, it is easy to leave a maintenance-mode service in place for years, quietly accumulating a dependency on something AWS has already decided to wind down.
The Real Story Is the AI Consolidation
Look again at Bucket 1 with the plumbing removed. Kendra was AWS's enterprise search service. Q Business was its enterprise AI assistant. Bedrock Agents just got "Classic" appended to its name, which is what a company does right before it ships the non-classic version. Add the row of SageMaker AI features and the shape is unmistakable: AWS is pruning its earlier, standalone AI and machine-learning products.
The important nuance, and the reason the honest framing is "consolidation" rather than "retreat": AWS is not stepping back from AI, it is concentrating it. The recommended path off Kendra is Amazon Bedrock Managed Knowledge Bases, not "go find another vendor." AWS is funnelling enterprise search, retrieval, and agent workloads onto Bedrock and away from the first-generation services that predated it.
For a customer, that reframes the decision. If you built on Kendra or Q Business, you did not pick a bad product. You picked the branch AWS has now stopped growing. The work ahead is not a fix, it is a move onto the platform AWS is actually investing in.
The One That Needs a Real Migration Project
Of everything in the announcement, Amazon WorkSpaces PCoIP is the item most likely to land on an ops team's plate as genuine work.
The verified timeline: PCoIP-based WorkSpaces Personal stops accepting new customers on July 31, 2026, and reaches end of support on October 31, 2027. You migrate to the Amazon DCV protocol, which is included at no additional cost, so pricing does not change. Through the October 2027 deadline, existing PCoIP WorkSpaces stay fully supported, with security patches and critical bug fixes maintained in coordination with HP.
That is roughly fifteen months of runway, which is generous, but it is a real project rather than a config flip. A protocol change touches the client software your users run, needs testing against your actual workloads, and benefits from a staged rollout. AWS has shipped tooling to help: a guided console action for the protocol switch, checkpoint snapshots so you can roll back, and session blocking during the migration. Start the assessment now while the runway is long, not in mid-2027 when it is not.
If You Built on Kendra or Q Business
AWS's recommended destination for Kendra is Amazon Bedrock Managed Knowledge Bases, which covers similar retrieval capabilities plus the generative and agentic features Kendra never had. The critical thing to understand is that this is a migration assessment, not an API-compatible swap. Treat it accordingly:
- Inventory your Kendra indexes, data-source connectors, custom metadata, and access-control behaviour.
- Rebuild an evaluation set from real production queries, not synthetic ones.
- Compare retrieval quality and, just as importantly, permission and ACL behaviour, because that is where RAG migrations quietly break.
- Recalculate cost and operational ownership under the new model before you touch any API.
The retrieval results and the security boundaries will both differ. Plan for a proper evaluation, not a lift-and-shift.
What To Do
- Check whether you run anything in the three buckets. The maintenance list is the easy one to miss precisely because nothing breaks and nothing pages you.
- Bucket 1 (maintenance): no scramble. Stop building anything new on these services and plan migration on your timeline. Note the one real near-term date inside it: IoT Device Defender Detect on August 31, 2026.
- Bucket 2 (sunset): these have deadlines. WorkSpaces PCoIP to DCV before October 31, 2027 is the big one, and fifteen months disappears faster than it sounds. Start now.
- Bucket 3 (already ended): if you are still on Chime SDK Carrier Voice Focus or SageMaker Ground Truth Plus, you are unsupported as of June 30, 2026. Move.
- If you are on Kendra or Q Business, begin the Bedrock Knowledge Bases assessment rather than assuming a drop-in replacement exists.
The measured version of this announcement is not "AWS is killing a dozen services, panic." It is AWS closing its first-generation AI and search products to new customers, consolidating that work onto Bedrock, and giving existing users a long runway on the handful that actually sunset. The whole task is knowing which bucket you are in, and refusing to keep building on a product AWS has already frozen.
If you want a fast audit of whether any of your accounts run services in these three buckets, and a migration sequenced by real deadline rather than by panic, that is exactly what our AWS cloud management and migration services work covers. For a related case on how AWS lifecycle language misleads, where the word "deprecation" hides both an already-passed risk and a delayed deadline, see our note on the AWS Lambda Node.js 20 timeline.
Sources
Talk to the engineer who will own your stack.
No account managers, no offshore handoff. Senior DevOps, direct. Tell us what you are dealing with and you get a straight answer.
Related News
The Node.js 20 Lambda Deadline Everyone Is Citing Is Wrong and the Real Risk Already Started
The AWS Lambda Node.js 20 deadline everyone is citing, August 31 and September 30, 2026, is out of date. AWS moved the block-create date to February 1, 2027 and block-update to March 3, 2027. But the number that matters is April 30, 2026, when the Node.js 20 runtime stopped getting language security patches. This covers what each date really means, the OS-versus-runtime patching nuance most guides miss, a one-command way to find every affected function, and the safe migration to nodejs22.x or nodejs24.x.
CloudAzure Returns 410 Gone for GPT-4o on October 1 and Auto-Upgrade Skips the Deployments That Matter
On October 1, 2026, Azure OpenAI in Microsoft Foundry retires the GA gpt-4o (2024-11-20) and gpt-4o-mini (2024-07-18) versions, after which calls to a retired deployment return HTTP 410 Gone. Standard-family deployments are auto-upgraded region by region, but Provisioned (PTU) deployments and anything set to NoAutoUpgrade are not, and Microsoft says the date is not extendable.
CloudStay on EKS 1.33 and AWS Starts Billing You 6x From August
Amazon EKS Kubernetes 1.33 leaves standard support on July 29, 2026, and any cluster still on it is enrolled in extended support by default at six times the standard control-plane rate. The hourly charge rises from 0.10 to 0.60 US dollars per cluster, and the standard patch cadence narrows. Fleets with many clusters feel it first.