Claude Sonnet 4 reached end of life on Amazon Bedrock on October 14, 2026, and requests to it now fail. Claude 3.x ends in more Regions on November 13, Opus 4.1 on January 8, the Llama, Mistral and DeepSeek batch on March 30, and Sonnet 4.5 on April 8. How to find what you call, the gotchas that bite, and how to move without a regression.
Claude Sonnet 4 (anthropic.claude-sonnet-4-20250514-v1:0) reached end of life on Amazon Bedrock on October 14, 2026, in every Region it was offered in. It had been Legacy since April 14 and in public extended access (higher prices, set by the provider) since July 14. Since the 14th, every request to it fails with:
Nothing migrates automatically. If a feature, a nightly job or an internal tool still pointed at that model ID — directly or through a us., eu., apac. or global. inference profile — it has been failing for ten days. Most teams find out from a customer, because a model error inside a background job rarely pages anybody.
From the Bedrock model lifecycle page, October 24, 2026. Lifecycle dates differ by Region — check yours.
Two things about that table matter more than the dates. First, the Regions column only lists where an end date has been announced: a model can still be Active in your Region today and gain an end date later. Second, Bedrock's dates are Bedrock's own. AWS is explicit that they can differ from the model provider's, and only the dates on the Bedrock pages apply to Bedrock traffic.
Book a free 30-minute call. You'll leave knowing what we'd do first and roughly what it costs, whether or not you hire us.
For models launched before September 7, 2026, a model is Legacy for at least six months. After at least three of those months it enters "public extended access", where the provider can raise the price for the remaining time. Opus 4.1 entered it on October 8. If your Bedrock bill went up this month with no change in traffic, check whether you are paying extended-access rates on a model you meant to have replaced.
Once a model is Legacy, new customers can't use it, and existing customers may lose access after 15 days of inactivity. A monthly report, a quarterly re-index or a rarely used admin tool can lose its model weeks before the published end date, simply because it didn't run in time. Jobs that run less often than every two weeks are the first thing to move.
When a base model goes Legacy you can no longer start fine-tuning jobs on it or create new Provisioned Throughput for it. Existing deployments keep working until end of life. If you depend on a customized model, its retraining on a current base model belongs in this quarter's plan, not next quarter's.
Your code is the least reliable place to look — model IDs live in environment variables, prompt-management configs, Knowledge Base settings, Step Functions definitions and other teams' Lambdas. Ask the platform instead:
CloudWatch: the AWS/Bedrock namespace reports Invocations by ModelId. Check every account and every Region you use, over at least 30 days so infrequent jobs show up.
CloudTrail: InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream events carry the model ID and the IAM principal that called it, which tells you whose code it is.
Model invocation logging, if enabled, gives the same per request, including inference profile IDs.
ListFoundationModels in each Region returns a modelLifecycle field per model — that is the authoritative state for your Region, and it is easy to check on a schedule.
Bedrock Knowledge Bases, Agents and Flows store a model ARN in their own configuration. Search those too; they don't appear in your application code at all.
Swapping the model string is the easy part. The newer models are better on average and different in the details: they format differently, follow instructions more literally, call tools in a different order and cost a different amount per task. A safe migration has four steps:
Build a small evaluation set from real traffic — fifty to two hundred representative requests with what a good answer looks like. Invocation logs or your LLM observability tool are where these come from.
Run the current model and the candidate side by side on that set, scoring with code checks where you can (valid JSON, right tool called, required fields present) and an LLM judge or a human where you can't.
Fix prompts for the differences you find, then re-run. Most regressions are one prompt instruction that the old model ignored and the new one follows exactly.
Ship behind a flag or a percentage rollout, compare cost and error rates for a week, then remove the old model ID everywhere — including the inference profiles and IAM policies that name it.
Sonnet 5.5 arrived on Bedrock on September 28, 2026, with a 1M-token context window, 128K output tokens and an end of life no sooner than September 28, 2027. From its model card, three details change code rather than just a model string:
Client-side tool calling supports tool_choice "auto" and "none" only. If you force a specific tool to get structured output, switch to Bedrock's structured outputs, which this model supports.
It is offered through cross-Region inference profiles (us., eu., au. and global.) rather than In-Region. In several APAC Regions only the global profile is available — if your data must stay in-country, check this before you plan the move.
Adaptive thinking is on by default at a "high" effort level. That is the right default for hard tasks and an unnecessary cost for simple classification; set the effort level per use case.
Models launched on or after September 7, 2026 follow a newer lifecycle policy: each model card states an "EOL no sooner than" date and a Legacy period, usually six months but 45 days for some. Record both for every model you adopt, because the next migration date is known the day you start.
Keep model IDs in one place per service — configuration, not code — so a migration is a config change plus an eval run.
Put a lifecycle check in CI or a scheduled job: call ListFoundationModels in each Region you use and fail loudly when anything you depend on is Legacy.
Keep the evaluation set. The first migration costs a week to build it; every one after that costs an afternoon.
Route through one gateway or shared client so a fallback model can be switched on in minutes when a deadline is missed.
Once a month: what changed in tech that month and what it means for you, from cloud and AI to security and the tools you build with. Every item links to its source. No sales pitch, unsubscribe in one click.
Book a free 30-minute call about it. Bring your setup and your bill; you'll leave knowing what we'd do first and roughly what it costs, whether or not you hire us.
Or keep reading
Generative AI for Law Firms: What ABA Opinion 512 Requires, and a Private Setup That Meets It
ABA Formal Opinion 512 applies existing duties to generative AI: competence, confidentiality, communication, candour and reasonable fees. Informed client consent is needed before putting client information into tools that learn from it, and boilerplate engagement language is not enough. The options for firms, from business chat plans to a private assistant on AWS, and what each means for those duties.
Read
AI Search Over Specs, RFIs and Drawings for Engineering Firms: What Works and What to Check First
Engineering firms sit on decades of specifications, RFIs, submittals, calculations and drawings that nobody can search well. Retrieval with a model on Bedrock makes them answerable, with citations. What it is good at, the licensing and client-data questions to settle first, and why the engineer of record stays responsible.
Read
As of October 24, 2026. Provided as is, for general information, and accurate to the best of our knowledge on the date shown. Not legal, compliance, tax, financial or other professional advice. Prices, limits and vendor services change; check current vendor documentation before relying on anything here. Terms of use.