The week infrastructure caught up to the models
Kimi K3 on Bedrock, SageMaker HyperPod Gateway and Astra for Law, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
Amazon Web Services
Kimi K3 on Bedrock
DEPLOYif you run long coding or knowledge workflows
WHAT SHIPPEDMoonshot AI's Kimi K3 is available on Amazon Bedrock, the first open model to reach 2.8 trillion parameters, with a 1-million-token context window and a 2.5x improvement in scaling efficiency over Kimi K2.
THE CATCHIt is the first open-weight model on Bedrock to support explicit prompt caching, so cached input tokens are billed at a discounted rate and do not count against input-tokens-per-minute quotas.
IF YOU RUNcoding agents or knowledge work that resends stable context like repository instructions or reference documents
SOURCEAmazon Web Services, Introducing Kimi K3 on Amazon Bedrock
Amazon Web Services
SageMaker HyperPod Gateway
DEPLOYif you run LLMs on Kubernetes GPU clusters
WHAT SHIPPEDAmazon SageMaker HyperPod Inference Gateway is a Kubernetes-native, GPU-aware routing system that deploys as a single EKS managed addon and reduces first-token latency by up to 82%.
THE CATCHThe gateway routes on real-time GPU signals including KV cache utilization, queue depth, LoRA adapter residency, and prefix cache hit rate, with no changes to model servers or client applications.
IF YOU RUNmixed GPU generations, bursty traffic, or shared prompt prefixes on vLLM or SGLang
SOURCEAmazon Web Services, Introducing Amazon SageMaker HyperPod Inference Gateway
OpenAI
Astra for Law
WATCHif you work in legal tech
WHAT SHIPPEDOpenAI introduced Astra for Law, combining GPT-6 Astra with a legal search index covering more than 230 million URLs and 26 partner-built plugins, passing the Vals AI Legal Research Bench overall correctness check on 54.0% of questions versus 38.7% for GPT-6 Astra with web search alone.
THE CATCHAstra for Law is initially offered only to selected law firms through Trusted Access in ChatGPT and Codex, and is coming soon to the API, so most teams cannot build on it yet.
IF YOU RUNlegal research or deal diligence and need answers grounded in citable authorities
SOURCEOpenAI, Introducing Astra for Law
The Deploy Log
Models and APIs
| DEPLOYif you run long-running or bursty agents AgentCore runtime. The new AgentCore runtime delivers a P75 cold start latency of about 2 seconds from a 200 MB image to 2 GB, while the original runtime rose from roughly 5.4 seconds to nearly 30 seconds. Amazon Web Services |
| DEPLOYif you run coding agents or tool loops GLM 5.3 FlashX. GLM 5.3 FlashX is available on AI Gateway, serving Z.ai's multimodal coding model at about 200 tokens per second with no markup and no platform fee on inference. Vercel |
| DEPLOYif you build voice applications GPT-Live 1. GPT-Live 1 from OpenAI is available on AI Gateway as a full-duplex voice model that can listen and speak at the same time, with client delegation to any text model on AI Gateway. Vercel |
Tools and products
| WATCHif you buy Apple hardware this cycle iPhone 18 Pro lineup. Apple debuted iPhone 18 Pro and iPhone 18 Pro Max, Apple Watch Series 12, Apple Watch Ultra 4, and AirPods 5 at Apple Store locations worldwide on Friday, September 18. Apple |
| DEPLOYif you study or train with purchased ebooks Expert Intelligence in Notebook. Google Workspace introduced Expert Intelligence in Gemini Notebook, featuring more than 100,000 books from major publishers that users can add as sources for grounded responses, infographics, audio overviews, and quizzes. Google Workspace |
| DEPLOYif you use Notebook for study Gemini Notebook learning tools. Gemini Notebook added an audio recorder, real-time conversation in nearly 100 languages, interactive learning overviews, new quiz formats, and short video overviews in more than 80 languages. Google Workspace |
| DEPLOYif you organize course or project materials Notebooks in Gemini. Notebooks in Gemini are now available to students, educators, and professionals, letting users add up to 10 sources to a notebook and combine the Gemini app with Gemini Notebook for study guides and video overviews. Google Workspace |
| DEPLOYif you run Actions on ubuntu-latest Ubuntu 26 runner image. The Ubuntu 26.04 runner image for GitHub Actions is now fully supported on x64 and arm64, and the ubuntu-latest label will migrate from Ubuntu 24.04 to Ubuntu 26.04 between October 19 and November 19, 2026. GitHub |
| DEPLOYif you run public repositories with Actions Workflow execution protections. Workflow execution protections for GitHub Actions are generally available with workflow file targeting, insights, and a REST API, and a default rule disables pull_request_target for public repositories, enforced on November 2, 2026. GitHub |
| DEPLOYif you run local models on macOS Ollama v0.34.2. Ollama v0.34.2 added first-run setup with sign-in or local options, an ollama://apps shortcut, and fixed excessive memory growth during long generations with MLX speculative decoding. Ollama |
| DEPLOYif you run Claude Code in production Claude Code v2.1.277. Claude Code v2.1.277 added AGENTS.md support, a CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY flag, and fixed sessions hanging with no result after internal errors, now reporting the error and exiting with code 1. Anthropic |
| DEPLOYif you use Codex in local TUI Codex rust-v0.155.0. Codex rust-v0.155.0 added experimental voice conversations with live transcripts, Touch ID verification for MCP requests, and AWS credential export commands for Amazon Bedrock. OpenAI |
| DEPLOYif you use Codex with non-OpenAI providers Codex rust-v0.155.1. Codex rust-v0.155.1 fixed new local TUI sessions leaving reasoning summaries disabled by default, which had caused request rejection by providers that do not support them. OpenAI |
| WATCHif you use NIXL transport vLLM v0.30.0rc2. vLLM v0.30.0rc2 tagged a bugfix to avoid receive reports for notification-only requests in NIXL. vLLM |
| DEPLOYif you use HITL tool calls LangChain 1.4.2. LangChain 1.4.2 fixed preserving model-generated tool calls in human-in-the-loop tool call edits and added a notice to ToolMessage. LangChain |
| DEPLOYif you use Zed with third-party models Zed v1.20.2. Zed v1.20.2 fixed payment errors from non-Zed model providers incorrectly prompting users to upgrade to Zed Pro, and fixed different snippet extensions for the same language canceling each other. Zed |
Robotics, hardware and chips
| WATCHif you operate in Singapore transport Waymo Singapore. Waymo announced a partnership with Singapore's Ministry of Transport and Land Transport Authority to launch a fully autonomous, all-electric commercial ride-hailing service via the Waymo app in 2028, after serving more than 20 million fully autonomous rides. Waymo |
Left out
- Workspace Studio flows: the publisher's page does not show it available yet
- Jev on AI Gateway: the publisher's page does not show it available yet
- Copilot weekly releases: it did not carry a number that moved
- Copilot code review: it did not carry a number that moved
- Life Sciences Verification Program: it did not carry a number that moved
The pattern
The week's releases split into two tracks that no single item states. On one side, Amazon Web Services, OpenAI, Google Workspace, and Vercel pushed model access and routing infrastructure: Kimi K3 on Bedrock with explicit prompt caching, SageMaker HyperPod Inference Gateway cutting first-token latency by up to 82%, Astra for Law passing the Vals AI Legal Research Bench on 54.0% of questions versus 38.7% for GPT-6 Astra with web search alone, AgentCore runtime holding P75 cold start at about 2 seconds from a 200 MB image to 2 GB, GLM 5.3 FlashX at about 200 tokens per second with no markup, and GPT-Live 1 as a full-duplex voice model. On the other side, Apple, Waymo, and GitHub shipped physical and workflow defaults: iPhone 18 Pro and iPhone 18 Pro Max in stores on Friday, September 18, Waymo Singapore launching a commercial ride-hailing service in 2028 after more than 20 million fully autonomous rides, and Ubuntu 26.04 migrating the ubuntu-latest label between October 19 and November 19, 2026 while a default rule disables pull_request_target for public repositories enforced on November 2, 2026.
The connective tissue is that infrastructure now carries the same version churn as models. Claude Code v2.1.277, Codex rust-v0.155.0 and rust-v0.155.1, vLLM v0.30.0rc2, LangChain 1.4.2, Zed v1.20.2, and Ollama v0.34.2 all shipped fixes for routing, caching, or session state, while Google Workspace added more than 100,000 books, an audio recorder, real-time conversation in nearly 100 languages, and video overviews in more than 80 languages to Gemini Notebook. The items together say that the bottleneck has moved from model capability to delivery: cached tokens, GPU signals, cold starts, and workflow protections are the features that determine whether a 2.8 trillion parameter model or a 2028 ride-hailing service reaches a user.