AWS AI Updates: November 3, 2026

1. Amazon Textract Ships an Updated Text Detection Model With 40+ Languages and Sturdier Table Extraction

Amazon Textract now serves an updated text detection model behind the DetectDocumentText and AnalyzeDocument APIs, with no changes to request or response schemas; existing applications are moved to it automatically. Coverage expands to more than 40 languages, including Arabic, Hebrew, Hindi, Chinese (Simplified and Traditional), Japanese, Korean, Russian, and Ukrainian, which AWS says enables processing of previously unsupported scripts, and AWS cites better long-word recognition and handwriting detection. Multilingual table extraction is described as more robust on merged cells, skewed or rotated tables, and partial or corrupted tables. The model is live in every Region where Textract runs and the announcement does not mention a way to pin the previous model, so pipelines with regression tests on extracted text may see output changes. In a separate Nov 2 announcement, Textract launched in Asia Pacific (Tokyo) and Europe (Zurich), bringing it to 18 Regions, but only the text detection and document analysis operations are available there at launch; AnalyzeExpense, AnalyzeID, AnalyzeLending, and Custom Queries adapters are not. Source Source

2. Amazon Connect Customer Flows Can Now Call MCP Tools Through AgentCore Gateway

Amazon Connect Customer added an External Tool flow block that invokes Model Context Protocol (MCP) tools from third-party applications connected through an Amazon Bedrock AgentCore Gateway, so contact flows can read from or write to systems such as a CRM or banking platform without custom Lambda functions, reusing the same tools an organization’s AI agents call. Setup registers an MCP server integration in the Connect Customer console that points at an AgentCore gateway, and the gateway ID must be added to the gateway’s inbound “Allowed audiences” list or tool invocations fail. According to the External Tool block documentation, the block works in all flow types across voice, chat, task, and email; synchronous calls time out after at most 8 seconds and asynchronous calls after 60 seconds, errors are retried up to three times, and results land in $.ExternalTool.ResultData with a 32 KB cap (larger synchronous responses take the Error branch), each invocation overwriting the previous result. A companion launch lets the Wait block hold a flow until an asynchronous External Tool call completes or a Connect Case field meets a condition, such as a refund status being updated. The External Tool block is available in US East (N. Virginia), US West (Oregon), Canada (Central), Europe (Frankfurt, London), and Asia Pacific (Seoul, Singapore, Sydney, Tokyo). Source Source

3. AWS Distills a GLM-5.3 Shopping Agent Into Qwen3.5-4B on SageMaker AI and Matches the Teacher

AWS published a sequence-level knowledge distillation (SeqKD) recipe on Amazon SageMaker AI that trains a small model to run a multi-step shopping agent from a frontier teacher’s trajectories. GLM-5.3 on Amazon Bedrock, with reasoning on and temperature 1.0, ran the ShoppingBench agent twice on each of 426 training problems over a catalog of about 2.7 million products; the benchmark’s scorer kept only solved runs, which expanded into about 2,600 per-step training rows with loss applied only to the assistant turn (reasoning plus tool call). Qwen3.5-4B was fine-tuned with serverless SFT in SageMaker AI model customization (LoRA rank 64, three epochs, 16,384-token maximum length) and served as a LoRA adapter inference component on top of the base model on a single G-family instance. On 324 held-out problems, average Absolute Success Rate rose from 0.297 for the base model to 0.541, against 0.540 for the teacher with greedy decoding and 0.546 for an unnamed medium-sized proprietary frontier model, from a single run that AWS says can vary by a few points. Training on tool calls without the teacher’s reasoning cut the student to 0.415, the student’s IFEval prompt-strict score fell from 0.819 to 0.741, and AWS did not publish per-task cost or latency figures. Source

4. Claude Haiku 5.5 Comes to Kiro in AWS GovCloud (US)

Kiro added Anthropic’s Claude Haiku 5.5 to its IDE and CLI in the AWS GovCloud (US) Regions, making all three Claude 5.5 models (Opus, Sonnet, and Haiku) available there. Haiku 5.5 uses adaptive thinking with five effort levels from low to max (medium by default), has a 1M-token context window, and is billed at a 0.06x credit multiplier, rising to 0.3x for requests over 100K tokens; AWS positions it for summarizing and compacting long context, classification and routing, bulk text processing, small edits across many files, and parallel review subagents alongside Opus 5.5 and Sonnet 5.5. Users must update and restart the IDE or CLI to see it in the model selector. The Kiro GovCloud documentation notes that requests from GovCloud (US-East) are processed by Amazon Bedrock in GovCloud (US-West), and that a model’s availability in Kiro does not by itself mean FedRAMP High or DoD IL4/5 authorization. Source