Pulse.
This wall is the Wire — speed and source-age currency. For research and policy with an analytical spine, read Deep. For a curated Batch-like package, open This Week.
Wire · news · Lead story
GPT-6 Improves Prompt Caching with New Tools and Controls
OpenAI has introduced an improved prompt caching system for the GPT-6 family of models, featuring higher cache hit rates, new diagnostic tools, and explicit controls to reduce latency and costs for developers.
Source — OpenAI News · Sep 22, 2026
Fig. 03 — the reading roomED 003
450 more · ED 003
ReAdapt Augments ReAct for Relational Social Reasoning in LLM Agents
A new paper introduces ReAdapt, an augmentation to the ReAct loop designed to improve how LLM agents handle relational social reasoning, addressing situations where content and relational cues diverge.
Read →MedGate-Fusion Integrates Clinical Narratives and Biomarkers for Stroke Risk
A new multi-modal gated architecture, MedGate-Fusion, combines transformer-based embeddings of first-encounter clinical narratives with ten physiological risk markers to stratify prospective stroke risk.
Read →SMTB Algorithm Accelerates Structure-Mapping and Improves Performance in Large Domains
A new algorithm, Structure-Mapping with Tight Bounds (SMTB), demonstrates a 5 to 15 times speed improvement over the Structure-Mapping Engine (SME) and a 50% improvement in finding mappings within large nested domains.
Read →GitHub Copilot App Integrates OpenTelemetry for Agent Monitoring
The GitHub Copilot app now supports OpenTelemetry configuration through enterprise-managed settings, allowing administrators to monitor Copilot agent activity.
Read →GitHub Copilot for JetBrains 1.18.0 Introduces AI-Assisted Tool Approvals and Enhanced Agent Controls
GitHub Copilot for JetBrains version 1.18.0 includes AI-assisted tool approvals, expanded control over agent conversations, and shared organizational skills and instructions.
Read →GitHub Copilot CLI Enhances C++ Code Intelligence with Whole Codebase Indexing
GitHub Copilot CLI now offers faster C++ code intelligence through the implementation of whole codebase indexing, which creates a persistent index of symbols across an entire C++ project.
Read →GPT-6 Sol and Luna Models Now Generally Available on Amazon Bedrock
Amazon Bedrock now offers GPT-6 Sol and GPT-6 Luna, providing additional options for integrating intelligence and efficiency into various workloads.
Read →Vercel AI SDK Adds Claude Opus 5.5 Support, Enhances File Handling, and Improves Harness Compatibility
The Vercel AI SDK has been updated to include support for Claude Opus 5.5 and to securely handle URL-backed language model file outputs, according to recent changelog entries.
Read →Claude Opus 5.5 Now Available in GitHub Copilot and on AWS
Anthropic's newest Opus model, Claude Opus 5.5, is now available in GitHub Copilot and on Amazon Bedrock and Claude Platform on AWS.
Read →NVIDIA Confidential Computing Enables Secure AI Inference with Minimal Performance Overhead
NVIDIA Confidential Computing (CC) allows secure execution of large language model inference on trusted hardware, with recent optimizations demonstrating over 96% retention of output-token throughput compared to non-confidential environments.
Read →Claude Opus 5.5 Now Available on AWS
Anthropic's Claude Opus 5.5, designed for agentic coding, knowledge work, and long-running tasks, is now accessible on Amazon Bedrock and the Claude Platform on AWS.
Read →NVIDIA Topograph Enables Topology-Aware Workload Scheduling
NVIDIA has released Topograph, an open-source toolkit that discovers and normalizes cluster topology from cloud APIs or on-premises fabric systems, publishing this data for workload managers to optimize GPU workload placement.
Read →Vercel AI SDK Adds Claude Opus 5.5 Support
The Vercel AI SDK now supports the Claude Opus 5.5 model, integrating it into the @ai-sdk/anthropic and @ai-sdk/gateway packages.
Read →OpenAI Outlines Priorities for Third-Party AI Safety Assessments
OpenAI has published its priorities and principles for independent third-party assessments of frontier AI models and their safeguards, emphasizing rigor, security, and independence.
Read →Claude Opus 5.5: Updates for Agentic Workflows
Anthropic has released Claude Opus 5.5, a model designed for long-running agentic coding and knowledge work, introducing several breaking changes for existing applications.
Read →Reactiv and Trane Leverage Amazon Bedrock AgentCore for Automation
Reactiv utilized Amazon Bedrock AgentCore to develop a multi-agent AI Scheduler, reducing Shopify merchant configuration time by 80% and accelerating production by 33%.
Read →Trane Technologies Accelerates Building Diagnostics with Amazon Bedrock AgentCore
Trane Technologies developed an AI-powered agentic solution on Amazon Bedrock AgentCore, reducing a 20-minute diagnostic workflow to 20 seconds.
Read →NVIDIA DLSS 5 Introduces 3D-Guided Neural Rendering for Game Developers
NVIDIA DLSS 5, featuring 3D-Guided Neural Rendering, is now available in NBA 2K27 for GeForce RTX 50 Series GPUs, allowing game developers to enhance lighting and material detail while preserving artistic intent.
Read →Foundry Hosted Agent Isolation with Microsoft Agent Framework
Microsoft's Agent Framework introduces Foundry hosted agent isolation, offering independent controls for user data access and the persistence of code and files.
Read →Tree-Structured Expression Serialization in Language Models Shows Lossy, Asymmetric Communication
A round-trip study on arXiv cs.AI reveals that serializing tree-structured expressions into natural language and then extracting them is a lossy and asymmetric communication channel for language models, with failures primarily originating during generation.
Read →GUARD Introduces Natural Forgetting for Large Reasoning Models
A new method, Guided Answer-Reasoning Distillation (GUARD), aims to enable large reasoning models to unlearn specific content by guiding them toward a non-disclosing chain-of-thought and a refusal-style answer.
Read →Biased AI Improves Human Performance While Reducing Perceived Helpfulness
A study involving 5,000 participants found that intentionally biased AI assistants improved human performance in tasks compared to neutral AI, despite participants undervaluing the biased systems.
Read →Vercel AI SDK for Svelte Updated to Version 5.0.109
Vercel has released version 5.0.109 of its AI SDK for Svelte, which includes a fix for handling replacement completion requests.
Read →Vercel AI SDK Updates Alibaba and xAI Integrations
The Vercel AI SDK has received updates for its Alibaba and xAI integrations, including support for model-specific structured output modes and the addition of the Grok 4.7 model ID.
Read →Vercel AI SDK Updates Alibaba and XAI Integrations, Adds Grok 4.7 Model ID
Vercel's AI SDK has released updates for its Alibaba and XAI integrations, including a fix for model-specific structured output modes and the addition of the Grok 4.7 model ID.
Read →Anthropic Introduces Claude Fable 5 and Claude Mythos 5 Models
Anthropic has released Claude Fable 5 and Claude Mythos 5, models designed for demanding reasoning and agentic work, with Claude Fable 5 including safety classifiers and Claude Mythos 5 available in limited release.
Read →Migrating to Claude Mythos 5 and Claude Fable 5
Anthropic has released a guide for migrating to Claude Mythos 5 and Claude Fable 5 from previous models, detailing API changes and providing migration checklists.
Read →Claude API Spend Limits
Anthropic's Claude API reference includes endpoints for managing and monitoring spend limits for individual members.
Read →Artifacts Endpoints in Claude API Reference
The Claude API reference now includes Artifacts endpoints, allowing users to retrieve artifact-creation activity for a specific day, categorized by MIME type.
Read →NVIDIA Dynamo-Triton Integrates TensorRT Multi-Device Inference
NVIDIA has introduced TensorRT multi-device inference, enabling a single TensorRT network to operate across multiple GPUs, supported from TensorRT 11.0 and integrated into NVIDIA Dynamo-Triton release 26.07.
Read →Evaluating AI Agents: From Tool Calls to Task Completion
NVIDIA details an agent evaluation framework that shifts focus from individual function calls to full task completion, using executable environments to track state across multi-step tool use.
Read →Vercel AI SDK Adds Grok 4.7 Model ID
The Vercel AI SDK has integrated support for the Grok 4.7 model ID in its @ai-sdk/xai package, according to recent changelog entries.
Read →Vercel AI SDK Adds Grok 4.7 Model ID
The Vercel AI SDK has updated its @ai-sdk/xai package to include support for the Grok 4.7 model ID, according to a recent changelog entry.
Read →Vercel AI SDK Updates Address Tool-Call Repairs and File Data Preservation
Vercel has released updates to its AI SDK, including version @ai-sdk/react@4.0.111 and @ai-sdk/rsc@3.0.108, which address issues with pending tool-call repairs and the preservation of file data when adapting language models.
Read →Vercel AI SDK Updates Address Tool-Call Repair and Language Model Adaptation
Vercel has released updates across its AI SDK packages, including fixes for pending tool-call repairs during generation cancellation and improvements to preserving file data when adapting v3 language models.
Read →Vercel AI SDK Updates Tool-Call Repair and Language Model Adaptation
Vercel has released updates to its AI SDK, addressing issues with pending tool-call repairs during generation cancellation and preserving file data when adapting v3 language models.
Read →Vercel AI SDK Updates: Tool-Call Repair Fix and File Data Preservation
Vercel's AI SDK, version ai@7.0.108, includes fixes for pending tool-call repairs and preservation of file data when adapting v3 language models.
Read →xAI's Grok 4.6 Arrives in Amazon Bedrock
xAI's Grok 4.6 is now accessible through Amazon Bedrock, offering a 500K token context window and four reasoning effort levels for various applications.
Read →EXL Medical IDP Solution on AWS Reduces Claims Review Time
EXL developed an AI-powered Medical intelligent document processing (IDP) solution on AWS, integrating IDP with domain-specific large language models on Amazon SageMaker and Amazon Bedrock to decrease medical claims review time.
Read →Meta Open-Sources Rebalancer for Assignment Problems
Meta has open-sourced Rebalancer, a library used for over nine years within the company to solve resource allocation and assignment problems by separating problem specification from solution.
Read →
Why an edition, not a feed
News should be curated like a gallery — not poured like a firehose.
Each monthly edition (ED 003 = September 2026) keeps what changes your decisions on the wall. Older editions stay forever — open any ED above to re-hang that month.