# AI 24 Updates · 2026-11-20-0800

Published: November 20, 2026 · 07:59 CET

## In Brief

- Tencent — Hy4 flagship is now open-weight, 770B MoE with 1M context
- OpenAI API — Reserved Tier for guaranteed capacity; Flex beta winding down
- Amazon Bedrock AgentCore — runtime instances now support nested virtualization
- Amazon SageMaker AI — detects slow responses, idle GPUs and missed optimizations on GenAI endpoints
- Amazon Bedrock AgentCore Identity — adds private certificate authority support
- Amazon Connect — AI post-contact summaries for chats in the Contact Control Panel
- AWS Revenue Measurement — now covers closed-source models in Amazon Bedrock
- AWS Partner Central Agents — now generate Value Realization Plans
- AWS Partner Central Agents — guide co-sell submissions and explain co-sell motions
- AWS Partner Central — AI-assisted listing experience for products, services and solutions
- Perplexity Decisions API — URL fetching and page freshness for web grounding
- Alibaba Cloud Model Studio — adds MindLab macaron-v1.1-venti-pro

---

## 1. Tencent — Hy4 flagship is now open-weight, 770B MoE with 1M context

Company: Tencent

Tencent released open weights for Hy4, its natively multimodal flagship MoE model, together with the Hy4 System Card; Together AI also serves the model.

**Facts:**
- Tencent released the open weights for Hy4, its natively multimodal flagship mixture-of-experts model, on Hugging Face, with code on GitHub.
- Hy4 has 770B total parameters, 49B active parameters and a 1M-token context window.
- Tencent also published the Hy4 System Card with a safety evaluation report; the weights come under the license in the tencent/Hy4 repository, and GGUF builds are on Hugging Face (AngelSlim/Hy4-GGUF).
- Together AI now serves Hy4 at $0.83 per 1M input tokens and $2.50 per 1M output tokens.
- Together AI describes Hy4 as the successor to Hy4 Preview, with visual understanding new to the Hy family and computer use through command-line tools and graphical interfaces.

**Sources:**
- [Hy4 weights on Hugging Face](https://huggingface.co/tencent/Hy4) (official)
- [Hy4 on GitHub](https://github.com/Tencent-Hunyuan/Hy4) (official)
- [Together AI model page](https://www.together.ai/models/tencent-hy4) (official)
- [Tencent Hy on X](https://x.com/TencentHunyuan/status/2123625549147070538) (official)
- [Tencent Hy on X: Together AI](https://x.com/TencentHunyuan/status/2123638726601621727) (official)

## 2. OpenAI API — Reserved Tier for guaranteed capacity; Flex beta winding down

Company: OpenAI

OpenAI's API docs add a Reserved Tier guide for Enterprise customers who reserve capacity for a specific model, and the pricing page says the Flex processing beta is being wound down.

**Facts:**
- OpenAI's API documentation now includes Reserved Tier, which gives Enterprise customers guaranteed API capacity for a specific model, up to the amount they reserve in dollars per minute.
- When demand is high, OpenAI prioritizes requests within a reservation over pay-as-you-go traffic; the reservation covers eligible Standard and Fast mode requests for that model.
- Customers pay for reserved capacity even when they use less; usage above it is billed at pay-as-you-go rates and normal rate limits.
- Reserved Tier is available on select Astra, Sol, Terra and Luna models through sales.
- The API pricing page now says Flex stays available on the listed models while OpenAI winds down the beta, Flex will not be added to new models, and a retirement date will be published in the Flex processing guide.

**Sources:**
- [OpenAI API docs: Reserved Tier](https://developers.openai.com/api/docs/guides/reserved-tier) (official)
- [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) (official)

## 3. Amazon Bedrock AgentCore — runtime instances now support nested virtualization

Company: Amazon

Agents on Amazon Bedrock AgentCore runtime instances can now run their own virtual machines and emulators with KVM inside a session.

**Facts:**
- Amazon Bedrock AgentCore runtime instances now support nested virtualization, so agents can run virtual machines and emulators inside a session with KVM.
- Nested virtualization is turned on when creating a Linux capacity provider, and agents in every session on its instances can use KVM.
- AWS gives examples such as running Android emulators to build and test mobile apps, or lightweight VMs to isolate code the agents run.
- It works on supported Intel-based instance types, including the C7i, C8i, M7i, M8i, R7i and R8i families.
- It is available in all AWS Regions where AgentCore runtime instances are available.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/agentcore-nested-virtualization/) (official)

## 4. Amazon SageMaker AI — detects slow responses, idle GPUs and missed optimizations on GenAI endpoints

Company: Amazon

SageMaker AI now monitors response time and GPU use of generative AI inference endpoints by default and notifies customers about slowdowns, idle GPUs and configuration gaps.

**Facts:**
- Amazon SageMaker AI now monitors response time and GPU use of generative AI inference endpoints by default.
- Findings include slower responses, unused GPU capacity and configuration gaps such as an older serving container, missing speculative decoding or missing GPU kernel tuning.
- Customers get one AWS Health notification per endpoint when a change holds across the past week or a configuration gap is found; AWS says it does not read customer requests or responses.
- Notifications link to SageMaker Studio, where Generative AI Inference Recommendations compare configurations on response time, traffic and cost per million tokens without touching the production endpoint.
- Monitoring can be turned off at any time.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/sagemaker-monitors-endpoints/) (official)

## 5. Amazon Bedrock AgentCore Identity — adds private certificate authority support

Company: Amazon

AgentCore Identity can now trust certificates from private CAs when agents connect to identity providers hosted in a VPC.

**Facts:**
- Amazon Bedrock AgentCore Identity now supports private certificate authorities for custom OAuth 2.0 credential providers.
- Connections to identity providers hosted in a VPC can trust certificates from AWS Private CA, enterprise CAs or self-signed root CAs.
- Previously the identity provider had to present a publicly trusted TLS certificate even over the private VPC Lattice connection.
- The CA certificate is stored in AWS Secrets Manager or Amazon S3 and can be rotated without downtime using a combined PEM file.
- It is available in all AWS Regions where AgentCore Identity supports private IdP connectivity.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/bedrock-agentcore-identity-private-ca/) (official)

## 6. Amazon Connect — AI post-contact summaries for chats in the Contact Control Panel

Company: Amazon

Amazon Connect Customer now shows agents a generative AI summary of each chat in the CCP seconds after the chat ends.

**Facts:**
- Amazon Connect Customer now gives agents a generative AI post-contact summary of each chat in the Contact Control Panel within seconds after the chat ends.
- The summary covers what the customer needed, the actions the agent took and any open items, and appears at the end of the transcript.
- Such summaries were already available for voice and email contacts; this launch adds chat.
- Summaries follow the flow's redaction settings and the same security profile permissions as voice summaries.
- It is available in 10 AWS Regions, including US East (N. Virginia), US West (Oregon), Europe (Frankfurt) and Europe (London).

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/connect-ai-chat-summaries-ccp/) (official)

## 7. AWS Revenue Measurement — now covers closed-source models in Amazon Bedrock

Company: Amazon

Partners now get attributed revenue for their solutions' use of closed-source models in Amazon Bedrock.

**Facts:**
- AWS Revenue Measurement, previously Partner Revenue Measurement, now covers closed-source models in Amazon Bedrock.
- Until now only Amazon models such as Amazon Nova and open-weight models in Bedrock were included.
- Attributed revenue for Amazon Bedrock now reflects total Bedrock charges, including closed-source model usage.
- Partners already using Resource Tagging, User Agent string, Identity Tagging or SaaS Account Tagging get the expanded coverage with no extra work.
- AWS Revenue Measurement is generally available in all commercial regions.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/revenue-measurement-bedrock/) (official)

## 8. AWS Partner Central Agents — now generate Value Realization Plans

Company: Amazon

AWS Partner Central Agents can now draft Value Realization Plans for the Business Value Realization motion.

**Facts:**
- AWS Partner Central Agents can now generate Value Realization Plans (VRPs) for the Business Value Realization (BVR) motion.
- A generated VRP includes the customer's business goals, a plan for adoption, change management and governance, and baseline KPIs for partners to review.
- The agents recommend KPIs, ramp targets and baseline measurements from nomination data, opportunity context and service benchmarks, and flag gaps before submission.
- BVR is open to consulting, system integrator and managed services partners with Advanced or Premier tier status and a qualifying domain competency.
- The capability is available in the US East (N. Virginia) Region.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/partner-central-agents-bvr-vrp) (official)

## 9. AWS Partner Central Agents — guide co-sell submissions and explain co-sell motions

Company: Amazon

AWS Partner Central Agents now guide partners to higher-quality opportunity submissions and explain why AWS assigned each co-sell motion.

**Facts:**
- AWS Partner Central Agents now guide partners to create and enrich opportunities for a higher-quality submission.
- The agents ask for details the Opportunity Quality Score measures, such as the customer's business problem, dated next steps and the AWS services the customer will adopt.
- After submission, the agents explain each opportunity's co-sell motion and why AWS assigned it, and recommend next steps.
- Partners can also use the agentic experience in AWS Quick and Kiro, or through MCP in their own CRM.
- It is available to AWS Partners in the US East (N. Virginia) Region.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/partner-central-agents-guide-high-quality-opp) (official)

## 10. AWS Partner Central — AI-assisted listing experience for products, services and solutions

Company: Amazon

AWS Marketplace partners can now draft and improve listings in AWS Partner Central with an AI-assisted listing experience.

**Facts:**
- AWS Partner Central now has an AI-assisted listing experience for AWS Marketplace partners.
- It drafts listing content from a partner's website and other materials and shows which sections need clearer details, use cases or evidence of customer outcomes.
- For single-AMI products it also drafts the usage instructions buyers see after subscribing.
- Partners can ask the assistant metering questions while integrating, and review all content before publishing.
- It is available in all AWS Regions except the AWS GovCloud (US) and China Regions.

**Sources:**
- [AWS What's New](https://aws.amazon.com/about-aws/whats-new/2026/11/partner-central-ai-listing/) (official)

## 11. Perplexity Decisions API — URL fetching and page freshness for web grounding

Company: Perplexity

Perplexity's Decisions API can now fetch a page from a URL before searching in web_grounding mode and refresh stored copies older than a set age.

**Facts:**
- In Perplexity's Decisions API, a page URL can now be passed as state with mode set to web_grounding to fetch its text before searching.
- The fetched text replaces the URL before search; the response returns the page in fetched_results and search evidence in search_results.
- A new max_staleness setting sets the maximum page age in seconds since the last crawl; an older stored copy triggers a fresh crawl.
- Each successful URL fetch costs $0.0005, on top of searches and input tokens.
- Web grounding itself runs a Fast Search for each question, up to 100 questions, at $0.001 per search.

**Sources:**
- [Perplexity API changelog](https://docs.perplexity.ai/changelog/changelog) (official)

## 12. Alibaba Cloud Model Studio — adds MindLab macaron-v1.1-venti-pro

Company: Alibaba (Qwen)

MindLab's macaron-v1.1-venti-pro is now available on Alibaba Cloud Model Studio in the Singapore region.

**Facts:**
- Alibaba Cloud Model Studio added mindlab/macaron-v1.1-venti-pro on November 20.
- The model is available in the Singapore region under the International deployment scope.
- Alibaba Cloud says it is built on MindLab's MoL architecture and supports a 1M context window and image understanding.
- It is listed with enhanced Chat, Agent, Coding and UI4A capabilities.
- It supports OpenAI-compatible and Anthropic-compatible APIs.

**Sources:**
- [Alibaba Cloud Model Studio release notes](https://www.alibabacloud.com/help/en/model-studio/newly-released-models) (official)
