Enterprise Claude on Google Cloud (Anthropic managed) · AWS Continuum GA · AgentCore Gateway fallback · Agent Compute Engine (preview soon) · NVIDIA Minerva
Enterprise this window: Claude Models on Google Cloud (Anthropic managed) (Claude release notes Oct 22; Google Cloud blog 5:00 PM ET). AWS Continuum for code vulnerabilities GA (News Blog 4:33 PM ET; What’s New 5:00 PM ET). AgentCore Gateway InvokeModel + fallback; Google Agent Compute Engine announced; NVIDIA Minerva.
Release notes Oct 22: launched Claude Models on Google Cloud (Anthropic managed); Google Cloud IAM controls access; Google Cloud Marketplace handles billing; fast mode for Claude Opus 5.5. Docs: “the Claude API, run by Anthropic and reached through your Google Cloud project”; setup “create[s] a new Anthropic organization for your Google Cloud project. It’s separate from any Anthropic organization your company already has. That includes a Claude Enterprise organization”; “Discounts can’t be applied to past usage”; three predefined IAM roles granted at project level (apply to every workspace — separate teams need separate projects); console sign-in through Google Cloud, not email. Google Cloud blog (Michael Gerstenhaber; RSS Oct 22 5:00 PM ET, page shows Oct 23): Anthropic-managed = day-zero models and Claude Platform features (Managed Agents beta, skills, code execution), Anthropic as data processor under Anthropic’s Commercial Terms, IAM + Cloud Audit Logging + commitment drawdown. Google-managed = Google as processor under existing agreements, VPC Service Controls, CMEK, US/EU/APAC residency, BigQuery and Model Armor integrations, features “roll out over time.” Same pricing. Models listed: Claude Fable 5.1, Sonnet 5.5, Haiku 5.5, Opus 5.5. Claude blog companion: “The full Claude Platform, available on Google Cloud.”
News Blog (Esra Kayabali) Oct 22 4:33 PM ET; What’s New 5:00 PM ET. Builds a context graph from code repositories (GitHub integration or an S3 bucket in the same account/Region) and the AWS environment (a collector gathers AWS Config resources and network configuration for reachability). Ingests existing findings, scans code, assigns an evidence level, runs reproducible exploits “in an isolated sandbox” with proof-of-concept code, gives attack path and blast radius, ranks by business impact, and opens fixes (patch, dependency upgrade, config change) as pull requests. Blog: “Every fix requires human approval … In the future, you will be able to … let Continuum apply those fixes automatically.” What’s New: “starts in Learn mode … you switch to the Enforce mode and let it act autonomous within the guardrails you set.” Mild contradiction: the blog presents autonomous fixes as future; What’s New describes an Enforce mode now. Pricing and Regions per AWS pages. Distinct from the Oct 5 Continuum penetration-testing preview.
AgentCore Gateway (What’s New Oct 22 4:00 AM ET): Bedrock InvokeModel APIs through the gateway plus model fallback routing. Google Agent Compute Engine (Google Cloud blog Oct 22 12:00 AM ET + 9:00 AM deep dive): microVM sandboxes for agents with snapshot/restore and short-lived credentials, plus expanded agent governance; preview “in the coming months” — Watch only. NVIDIA Minerva (Oct 22 11:00 AM ET): AIOps/observability platform for AI factories (title-level). Title-level reads, table rows only.
Google Cloud (Oct 21, updated 3:30 and 4:25 PM PDT): CVE.org assigned both Oct 21; CVSS 9.2 Critical each. 109618: out-of-memory interrupts PHP while request data is partially initialized; cleanup treats it as valid. 109619: request-owned state survives after its memory is freed; a later request on the same worker uses stale state. Affected 5.4 through 8.5.11; fixed 8.5.12, 8.4.27, 8.3.36, 8.2.35; no fixes for 8.1 or earlier. App Engine standard (8.2+) auto-patched; App Engine flex must redeploy; Cloud Run / Functions auto-patched only with automatic base-image updates; self-built images are your job; Cloud Armor managed rules mitigate. Credits Wiz (“a mix of models”) and says Gemini 4 Argon contributed to finding, exploit development, detection, and patch validation. Cloudflare (Oct 21 9:56 AM ET): disclosed by Calif.io, Wiz, OpenAI, Quarkslab; WAF protections developed since early October, live for Free and paid WAF from Oct 21 03:02 UTC (~9 h before disclosure); “the vulnerability was leaked before the official release date and we observed attack traffic”; new full-transaction (request + response) detection; Managed Free Ruleset and Managed Ruleset rule IDs listed; paid plans should confirm block mode. Cloudflare’s suggested agent prompt notes distro packages backport without version bumps (check changelogs for the CVE IDs), remi repo not same-day, restart every worker, and interim php.ini mitigations (not a fix). OpenAI (Oct 21): Calif found a flaw using GPT-5.6 Sol under Patch the Planet; Wiz independently found and reported both; PHP maintainers led fixes; Codex skill PHP version checker; “frontier AI found these vulnerabilities before attackers did” (contradicted by Cloudflare’s leak/attack-traffic report); five-stage coordinated-response framework; $1B subsidized Daybreak access for Frontline Defenders; says it engaged ONCD, CISA and a White House “Super Intelligence Force” (OpenAI’s description; no White House primary found). EU (Oct 21): joint statement by DG CNECT, ENISA, CERT-EU, Europol EC3, CSIRTs network advising immediate action. CISA KEV catalog 2026.10.21 did not list either CVE at our check.
AWS What’s New Oct 21 11:00 AM ET: “Beam model by Reflection is now available on Amazon SageMaker JumpStart and AWS Trainium” (Beam was Issue 41’s lead); “AWS Neuron 2.33 adds native PyTorch support, Neuron Observer, and other new features” (follows TorchNeuron public beta, Issue 42). Also Oct 21: SageMaker AI serverless customization custom sequence lengths. Title-level reads.
Google Cloud blog Oct 21, 8:00 AM ET (Michael Gerstenhaber). Models: GPT-Astra, GPT-6.1 Sol, GPT-6 Luna — “rolling out now in private preview to select enterprise customers,” interest form. Transacted via Google Cloud Marketplace: single invoice, draws down commitments, counts toward Flexible Savings Plans. “Starting today” APIs + Codex CLI and Desktop can use the Gemini platform for inference. Governance claims: prompts/outputs stay in your perimeter, not used for training; ZDR “coming soon”; dedicated infrastructure with Confidential Compute “coming soon”; VPC Service Controls, unified audit logging. Coming soon: Gemini for business and Google Antigravity. “Only hyperscale cloud provider offering access to all three leading closed source LLMs” is Google’s claim. No matching OpenAI post in openai.com news RSS at our check.
Cloudflare blog Oct 20 22:20Z (≈6:20 PM ET). Timeline (UTC): Oct 19 ~17:00 aware; Oct 20 02:00 Cloudflare One Client / WARP updates (Cloudflare One Agent on iOS pending app review); 09:15 cloudflared update; 19:20 CDN + server-side Zero Trust mitigations fully deployed. Weak keys for eight widely trusted intermediates → forged certificates for any domain. No active exploitation identified on Cloudflare services. Customer actions: update WARP/1.1.1.1 and cloudflared, review mTLS trust stores / uploaded CAs, origin certificate configs, container images (Cloudflare Containers/Sandboxes); route container egress through Cloudflare to inherit its trust-store mitigation. “Updating the SDK alone may not change the underlying policy.” Do not resolve new errors by disabling verification. Cloudflare links a community-maintained patch list (not a primary; not cited here). This addresses certificate authentication, not TLS encryption.
NVIDIA newsroom Oct 21 (≈5:48 AM ET). Extends the open-source OpenShell secure agent runtime (Issue 26) into a full-stack reference: Vera Rubin for inference, Vera CPUs for agent harnesses/tool use, BlueField-4 DPUs running NVIDIA Sentry for out-of-band monitoring and “in-silicon” enforcement (“quarantines and stops it in milliseconds”; “boosts agent capacity by up to 20%” — vendor-reported). Runs on Canonical, Red Hat, Windows with Microsoft Execution Containers; extends to RTX Spark, DGX Spark, DGX Station. Long partner list (Anthropic, AWS, Google Cloud, Microsoft, OpenAI, ServiceNow Arc GA, SAP Joule Studio, CrowdStrike, Palo Alto). Table row only.
Databricks blog Oct 21 (≈7:20 AM ET): AI Runtime (AIR) GA — managed GPU training inside Databricks with Unity Catalog governance, BYO code/containers/Ray, “seconds-level launch latency on RDMA-enabled GPUs” (vendor). Companion post with NVIDIA (7:30 AM ET). Oct 20: Agent Memory on Databricks + MemCurrent design post (title-level read). Table row only.
AWS What’s New Oct 20: Structured outputs for Claude models on Amazon Bedrock (10:00 AM ET); Bedrock Guardrails Standard tier for contextual grounding checks; Bedrock Advanced Prompt Optimization prompt shortening. Google Cloud Oct 21: Confidential G4 (NVIDIA Blackwell) GA; Vera Rubin GA with NVIDIA. Google Cloud Oct 20: Claude Opus 5.5 / Sonnet 5.5 in Antigravity. Lightning AI incident page (Oct 16) re-read Oct 21: platform “will remain inaccessible.” Title-level reads unless noted.
llm-d v1.0.0 GitHub release Oct 20 02:30Z (≈10:30 PM ET Oct 19); Google posts 11:30Z (7:30 AM ET Oct 20). CNCF Sandbox; co-founded by Google with Red Hat, IBM and others; 500+ contributors across 25+ orgs. 1.0 adds: priority/fairness flow control; v1 APIs; router high availability and zero-downtime, prefill/decode-aware rolling updates (LeaderWorkerSet); session-aware routing, turn-priority scheduling, native Responses API state; async processor pulling from Pub/Sub or Redis (llm-d-async); RL time-slicing (llm-d-rl-time-slicing). KV offload via vLLM OffloadingConnector, Mooncake, or SGLang HiCache to host RAM, then SSD or Managed Lustre. On GKE, Inference Gateway runs the router as a managed service. Google-reported: Ironwood agentic coding — median TTFT −85%+, throughput 2.4× vs a plain Kubernetes Service; Gemini platform prefix-cache hit 35% → 70%, TTFT −35%+ for coding agents; 17,000-node multi-region test 2.9× throughput at 99.9% success; A3 Mega 70B with Lustre tier ≈40% better TTFT, 30% lower e2e latency, ≈60% fewer GPU-hours at 50k-token prompts; RL duty cycle ≈40% → >70%. Customer quotes (Snap 75–80% prefix hit; ZenteiQ 1.47B → 5.97B tokens/day per 1,000 TPU v6e) are customer-reported. TorchTPU (google-pytorch/torch_tpu): co-developed with Meta’s PyTorch team; eager mode + torch.compile; vLLM backend with Inferact (vllm-torchtpu), SGLang backend with RadixArk (sglang-torchtpu); Transformers v5 and Diffusers run on TPU; TorchTitan recipes validated to 8,000+ Ironwood chips (up to 52% MFU on Qwen 3 8B, Google-reported); VeRL TPU backend + GRPO recipe (preview); Ray support; LiteRT joins the PyTorch Ecosystem; TPU agent suite preview.
Incident update on the Lightning AI blog, dated Oct 16 (lightning.ai/blog returned 404 at Issue 41’s check; reachable today). Identified Oct 15 23:00 UTC (7:00 PM ET) — unauthorized access to credentials on the Lightning Platform; all user access revoked Oct 16 01:00 UTC (9:00 PM ET Oct 15). Scope: Lightning Studio Platform services (Studio, Deployments, Pipelines, Drive); not GPU cluster rental, managed Kubernetes, or SLURM. Lightning believes certain customer credentials were used to access environments and connected third-party services; identified credentials revoked; external cybersecurity firm engaged; potentially affected customers notified. Recommended for all users: rotate secrets/API keys and revoke credentials stored in Studios (e.g. GitHub); update credentials for third-party services accessed from VMs; review third-party access logs. Lightning revoked identified keys for cloud accounts used on the platform. Page at our check (≈8:20 AM ET Oct 20) still says the platform “will remain inaccessible to all users.” PyTorch Lightning (the OSS library, 2.6.6) is not named as affected.
Posted Oct 19 19:12Z (≈3:12 PM ET). Monthly cost per review rose $1.22 (Mar) → $1.44 (Jun) → $1.74 (Jul) → $2.32 (Aug); early-August switch of all reviewers to Anthropic models raised cost per job $1.38 → $2.03 (code-quality reviewer on Claude Opus 4.7: ≈$0.25 → $0.88). Cache hit rate 85.7% (Apr) → 97% (Sep) while tokens/review tripled 0.91M → 2.99M. Spend split before work: output 23%, cache writes 43%, cache reads 34%; code-quality reviewer ≈16.6 calls/review, 72k tokens written to cache, 1.08M read. Root cause: OpenCode v1 built one system block (agent prompt + environment: working dir, date) with the only cache breakpoint at its end, so the 11k–21k-token shared prefix never cached across repos — first-request hit 14–26%. Sep 16: marker + plugin hook split the block; skills list disabled; tool descriptions frozen. Standards reviewer index moved into prompt: calls 15.9 → 9.6, $1.12 → $0.72. Sep 21/23: sub-reviewers stopped seeing the final-comment tool (they tried it in 57–76% of sessions). Result: $0.77 average (code quality at high effort), −68% vs $2.43 (two most expensive weeks); median 1m32s vs 4m09s; ~40% of savings from reviewer code, rest from model/effort choices. v2 shadow mode agreed with v1 on 98% of approvals. Flag: the post uses several baselines ($2.46 end-Aug, $2.32 Aug average, $2.43 two-week, $2.47 pre-change); all Cloudflare-reported.
Posted Oct 19 17:48Z (≈1:48 PM ET). Preview on Workers AI as @cf/cloudflare/clairic; “we intend to release the model open-weights soon.” Single forward pass, no output tokens, billed on input only, typically <1 s. Input: text plus up to four optional images and a map of typed fields (name, type, optional description). Returns spans with character offsets and confidence; typed numbers/dates/times/durations/lists; estimates as distributions within declared bounds (median + 80% interval, 99 percentiles on request); per-step forecasts; null when not found. Unlimited fields (split into parallel passes that each see the whole input). Sits beside Cloudflare’s Clef decision model and TypeSafe’s Jev in the “System One” framing. Accuracy, latency, and cost advantages over LLMs are Cloudflare-reported.
AWS What’s New Oct 19 13:00Z (9:00 AM ET): Grok 4.7 available on Amazon Bedrock in AWS GovCloud (US). AWS ML blog Oct 19 (6:32 PM ET): “Structured outputs for Anthropic Claude models on Amazon Bedrock” (not reviewed in depth). Google Cloud Oct 20 10:00Z (6:00 AM ET): Gemini for government services, Public Sector Summit highlights, Gemini on GDC enterprise search (sovereign RAG) — sales/availability posts, no operator delta pulled into the issue.
Google Cloud blog Oct 16, 2026. Recipes on AI Hypercomputer: A4X Max (GB300 NVL72) “up to 5x higher throughput and 8x lower latency” and G4 (RTX PRO 6000) “up to 2x higher throughput and 8x lower latency” Google/Reflection-reported, baseline unstated in body. Stack: GKE + llm-d EPP / Dynamo router for prefix- and load-aware routing so later agent turns hit the replica with the warm KV prefix; prefill/decode disaggregation; speculative decoding; hierarchical KV offload to host DRAM and Local SSD. Reflection quote repeats the 5x / 8x figures.
AWS ML blog Oct 16 15:23Z (≈11:23 AM ET) + What’s New (Oct 16): Beam on SageMaker JumpStart real-time endpoints in US East (N. Virginia) and US East (Ohio); Kubernetes path with JumpStart, SGLang, or vLLM; AWS Trainium listed. Same day: NVIDIA Nemotron 3.5 Super VL on JumpStart (What’s New). AWS ML blog also posted an AgentCore evaluation-framework guide (Oct 16).
Cloudflare blog Oct 16, 2026. Verified-bot status now comes with ongoing behavior evaluation: Cloudflare scores whether a verified crawler respects site preferences (robots directives, rate requests, opt-outs) and reports deviations, beyond identity verification. Operator hook: if you run an AI crawler or agent fetcher under Verified bot status, its behavior is now evaluated continuously, not just at verification; the post notes Cloudflare already removed one bot from the list in 2025 for trying to circumvent blocks.
Google Developers Blog Oct 15, 2026. Refresh tokens are sender-constrained to a client-generated asymmetric keypair via DPoP proof JWTs (htm, htu, iat, jti, nonce handling); access tokens (≈1 h) remain bearer to avoid per-request signature verification on high-QPS APIs. Application-layer, so it survives proxies that break mTLS (RFC 8705). Google Workspace CAA uses DPoP to validate agent identities at the Gemini Enterprise Agent Platform Agent Gateway. Multi-agent chains: delegate through STS token exchange so each agent gets a short-lived federated token bound to its own key (no key sharing). Recommended for BFF / token-mediating backend patterns. Does not protect browser session cookies (see DBSC work).
Cloudflare blog Oct 15, 2026. New Gateway action and custom validation functions (all Zero Trust plans) run a Cloudflare Worker inline during policy evaluation to transform traffic or bring your own data to DLP decisions (e.g., read JWT claims and inject headers instead of a sidecar NGINX pool).
AWS What’s New Oct 15, 2026: Muse Spark Open available in SageMaker JumpStart (ML blog post same day); Web Search on Bedrock AgentCore adds US East (Ohio), US West (Oregon), Asia Pacific (Sydney), and AWS GovCloud.
Anthropic news Oct 15, 2026 (updated same day re SecureBio terms). Six METR staff onboarded with company hardware, office space, and internal access roughly on par with Anthropic’s risk teams; exclusions include customer data, privileged material, and model architecture details. METR publishes risk analyses; Anthropic cannot veto, only require redaction in predefined categories, and METR may characterize redactions. Three evaluator access tiers (approval-gated publication → publish without approval → embedded). Procurement hook: ask vendors who outside the company can assess and publish on their risk posture.
Post Oct 14, 2026; Apigee hybrid 1.17.0 released Sep 14. OpenAPI 3.0.x → MCP Discovery Proxy; each operation becomes a tool (tools/list, tools/call over JSON-RPC). MCP data plane runs in your Kubernetes cluster; a sidecar pulls config from the control plane (no inbound path), reusing the apigee-watcher identity. Tool entitlement via API products (filters discovery; optional per-tool quota). Enable with enableMcpServer: true in overrides.yaml; upgrade apigee-operator chart before apigee-org (wrong order = no MCP pods); triggers staged canary of runtime pods; defaults 2 replicas → 10. Extracts W3C trace context from MCP _meta and forwards as traceparent/tracestate (SEP-2028 proposal). Subscription, Evaluation, and Data Residency orgs.
Announced Oct 14, 2026. Public Preview on Azure, Private Preview on AWS; Google Cloud in development. Delegate one subnet; Databricks runs a managed encrypted gateway configured in the Network Connectivity Configuration (NCC). Modes: specific_destinations (listed FQDNs only) and all_traffic (except more-specific Private Link/service-endpoint paths). Use cases named: Lakeflow Connect to private DBs, agents calling a private MCP server, private PyPI/Artifactory/Nexus, egress through your firewall, customer-owned source IP. Serverless Private Link stays the path for cloud-managed services (S3, ADLS). Databricks says >75% of interviewed customers cited networking as the blocker to serverless (vendor survey).
Published Oct 15, 8:00 AM ET. BYOC (AWS, Google Cloud, Azure) now supports on-demand capacity: indexes share one pool with dynamic sharding (add shards without rebuild or data movement); millions of namespaces per pool; hot namespaces can move to dedicated read nodes in the same cluster. Zero-access model kept (no SSH/VPN/inbound/standing cross-account role). Available to all BYOC customers.
Posted Oct 13, 2026. AWS AI/ML skill is now a core skill in the Agent Toolkit for AWS (supported path replacing prior AWS Labs skills/plugins). Pre-installed in SageMaker Studio JupyterLab. Capabilities: inference optimization; model customization / fine-tuning; serverful training; MLflow; JumpStart deploy/recommendations; processing jobs. No additional skill charge; pay for AWS resources used.
Posted Oct 12, 2026. STM priced by ingest $1.00/GB, retrieve $0.20/GB, storage $0.10/GB-month. Ingest/retrieve billed per event with 12 KB minimum and 64 KB maximum billable size; storage on actual event size. Replaces flat rate per event. Long-term memory pricing unchanged. Available in all Regions where AgentCore Memory is offered. AWS: “Most workloads see lower costs” — AWS claim. Operator lever: TTL on memory resources.
Published Oct 12, 2026. NVIDIA contributed Vera Rubin rack/compute tray designs via MGX; MGX 1.2 notes include rack-scale liquid cooling and liquid-cooled 5,000A busbar, racks up to 250 kW (vendor). Open-source DSX components: Config Manager, Cluster Readiness Engine, Rack Management Service, Cluster API Provider for NVIDIA Infrastructure Controller. Open Data Center Specification v0.7 (AMD, Google, Meta, NVIDIA). OCP Grid Resilience Specification (Google, Meta, Microsoft, NVIDIA): stay connected, restore near-full load in <2 seconds (spec claim). Meta Spectrum-6 / N-IceCube; Falcon Transport on future ConnectX-10 and BlueField with Google. Flag all performance/power claims as vendor/partner.
AWS What’s New Oct 9, 2026 17:44Z: Bedrock supports the reasoning.summary parameter for OpenAI models via the Responses API. Returns a human-readable summary in the summary array of the reasoning output item alongside the answer. Available for all OpenAI models on Bedrock wherever those GPT models are offered, including in-Region, GEO, and global cross-Region inference. Points to OpenAI’s reasoning-summaries docs for behavior details. No speed or price figures in the What’s New post.
AWS What’s New Oct 9, 2026 05:32Z: Pegasus 1.5 video-to-text model generates structured, time-coded metadata from an entire video in a single pass (vision, audio, on-screen text). Positioned for editorial segments, speaker changes, sports plays, brand appearances without a separate pre-index pipeline. Accessible via US and Global cross-Region inference profiles. Regions/APIs/pricing in Bedrock docs.
Cloudflare Blog Oct 9, 2026: Deno team joining Cloudflare; plan to merge celld ideas into open-source workerd so the Workers programming model (including Durable Objects) is first-class for self-hosting outside Cloudflare’s network. Ryan Dahl / Kenton Varda posts. Forward-looking on the merged effort (“coming months”); celld and workerd usable today. Operator relevance: escape hatch / self-host path for agent backends built on Workers primitives.
AWS What’s New 2026-10-08T18:51Z ≈ 2:51 PM ET. New cost dimension breaks Bedrock spend down by model (e.g. Claude Sonnet 5, Claude Haiku 4.5), model provider, inference type (input/output tokens) and feature (on-demand, reranker). Group and filter in Cost Explorer and Dashboards; filter Budgets. Combine with cost allocation tags on application inference profiles, projects or workspaces, and IAM principal tags. No additional charge; all Regions except GovCloud (US) and China (Beijing/Ningxia). Same afternoon (19:41Z): Ultrafast mode for OpenAI GPT-6.1 Sol on Bedrock (see Frontier).
PyTorch blog 2026-10-08T18:13Z. Dynamo unifies router, engine and KV-manager work around a session ID (plus parent_session_id for subagents). It maps headers coding agents already send — x-claude-code-session-id / x-claude-code-agent-id / x-claude-code-parent-agent-id, Codex and OpenCode equivalents — with no configuration. Covers session-linked tracing and replay, session-aware admission with backpressure at tool boundaries, an experimental shared-pool KV indexer, and a proposed KvHint interface, across vLLM and SGLang. No headline speedup quoted in the sections we checked.
Databricks blog 2026-10-08T22:00Z. Lakebase Postgres branches an entire database “in under a second, regardless of its size” with copy-on-write; branches scale to zero when idle. The post walks through one branch per agent for migrations and tests, retired when done, with GitHub Actions examples, and recommends branching from a seeded database rather than production to avoid exposing PII. How-to, not a new GA.
Keynote post Oct 8 (blog.google 12:05Z ≈ 8:05 AM ET). One agent and “a single API” for chat, autonomous objectives and code; headless operation; access from web, iOS, Android, Windows, Mac, CLI, Workspace, Microsoft 365 and Slack. Spawns temporary sub-agents with their own identities; “coworker agents” get their own Workspace account (@agents.company.com email, calendar, Drive). Model choice per job across the Gemini family and Anthropic’s Claude today, other models “in the future”. Tools registry (incl. any MCP server, on or off network), skills registry, four memory types (session, semantic, procedural, episodic). Governance: per-agent cryptographically attested identity stamped into logs and VMs, OAuth propagation, audit attributed to the agent, Agent Sandbox with its own network boundary, and Agent Gateway (“an AI network firewall”) for org-wide policy. Cost: multi-model orchestration, Smart Routing, real-time spend caps in Cloud Billing that pause the project’s agent. Industry specializations are in preview for Financial Services and Legal. The post gives no GA date or pricing for the core agent. “TPU 8i 80% better price-performance” and “per-token prices have dropped 98% since 2024” are Google claims; customer results are customer-reported via Google.
AWS What’s New Oct 7 (14:00Z ≈ 10 AM ET). On Amazon Bedrock via US, EU, AU and JP Geo cross-Region inference profiles plus the Global profile (e.g. global.anthropic.claude-haiku-5-5); in AWS GovCloud (US) on both bedrock-runtime and bedrock-mantle. Also via Claude Platform on AWS in North America. AWS repeats Anthropic’s “around 75% less than Claude Haiku 4.5” (Anthropic-reported). See the Bedrock docs for the Region list.
Databricks blog Oct 7: with OBO, an app calls supported Databricks APIs as the signed-in user, so Unity Catalog applies that user’s existing permissions (including row filters) without re-implementing governance in app code. Use app authorization for app-owned work and OBO for user-governed actions.
Cloudflare blog Oct 7: the Managed Defense harness groups related alerts, gathers evidence and flags missing sources. Initial analysis and scoring use Clef, Cloudflare’s open-source decision model; deeper analysis uses approved OpenAI Daybreak and Anthropic models (GPT-5.6 Cyber, Mythos). Architecture write-up; no new product SKU.
AWS What’s New 2026-10-05T21:42Z ≈ 5:42 PM ET; AWS ML blog the same day. 753B-parameter MoE with ~40B active per token; 1M-token context; up to 128K output; reasoning always on with selectable effort levels. Available to eligible enterprise customers. Cross-Region inference profiles us.zai.glm-5.3 and global.zai.glm-5.3. Explicit and implicit prompt caching, with explicit cache breakpoints needing ≥ 1,024 tokens. OpenAI-compatible Responses and Chat Completions plus InvokeModel / Converse. Flex, Priority, and Standard service tiers. Benchmarks such as CyberGym 84.5 and “50% over GLM 5.2” on an internal coding benchmark are Z.ai-reported. The blog demos the open-source Strix pentest agent and notes LiteLLM does not yet resolve bedrock/global.zai.glm-5.3. Run dual-use security demos only against systems you are authorized to test.
AWS What’s New 2026-10-05T08:10Z ≈ 4:10 AM ET. Speech-to-speech model on Bedrock. AWS claims improved reasoning, instruction following, and tool-calling accuracy and lower latency, with no figures published (AWS claim). Seven languages, voice and text in one session, asynchronous tool calling, 256K context. Strands Bidi Agents GA. Regions: US East (N. Virginia), US West (Oregon), Europe (Stockholm), Asia Pacific (Tokyo). Same pricing as Nova 2 Sonic.
AWS What’s New 2026-10-05T18:32Z ≈ 2:32 PM ET. Continuum for Penetration Testing (formerly AWS Security Agent) runs as a pipeline step. Findings land in pipeline output with severity, affected endpoints, and remediation, and re-test runs automatically after fixes. AWS says the auto-generated snippet takes “under five minutes” to set up, with no prior full penetration test required (AWS claim). Public preview.
AWS What’s New Oct 2, 2026. Amazon Bedrock AgentCore Gateway accepts TLS certificates signed by private CAs on MCP server, OpenAPI, and HTTP proxy (passthrough) targets. Register a private CA with targets that use private endpoints powered by Amazon VPC Lattice; gateway fetches the PEM-encoded CA from S3 or Secrets Manager as the outbound trust anchor — no intermediate ALB required. Available in all Regions where both AgentCore Gateway and VPC Lattice exist.
AWS What’s New Oct 2, 2026. Managed MCP server (part of Agent Toolkit for AWS) now also in Singapore, Sydney, Tokyo, Ireland, London and Oregon; full list: us-east-1, us-west-2, ap-southeast-1, ap-southeast-2, ap-northeast-1, eu-central-1, eu-west-1, eu-west-2. Server can still access services in all commercial Regions; regional endpoints keep request data in-Region for residency.
AWS What’s New Oct 1, 2026 (not in Issue 30). Preview; positioned as next-gen of Trusted Advisor + Well-Architected Tool. Correlates metrics and topology against Well-Architected practices; reviews Terraform, CDK and CloudFormation and returns IaC changes; recommendations ship with SSM runbooks / CLI scripts where applicable. Access in us-east-1, us-east-2, us-west-2 (onboard workloads from any commercial Region); requires an AWS Support plan.
What’s New Tue Sep 30, 2026. AWS announces UltraFast mode for GPT-6 Astra on Amazon Bedrock — premium speed tier for latency-sensitive apps (real-time coding assistants, interactive agents, customer-facing experiences). AWS cites OpenAI: up to 6× faster inference in the API, up to ~300 tokens/second — vendor / OpenAI-attributed. Start via Bedrock console or supported Bedrock APIs; Regions/endpoints/inference profiles/pricing → Bedrock docs. Distinct from Issue 28’s OpenAI API Ultrafast docs (service_tier: ultrafast). Do not claim Sol Ultrafast GA. Do not re-lead BMA (Issue 28).
What’s New Tue Sep 30, 2026. In-region support for Claude Opus 5.5 and Claude Sonnet 5 on Bedrock in UK (London) / eu-west-2. Aimed at FS / healthcare / public sector needing UK data processing. Inference on bedrock-runtime in London; AWS states processing never leaves the Region called. Do not conflate Sonnet 5 with Sonnet 5.5 (Issue 27) or Opus 5 with Opus 5.5.
Databricks blog Tue Sep 30, 2026. New AI Function ai_decide: fast structured decisions over governed data (routing, tagging, agent evals). Returns probability / named choice / ordered-scale score per question; positioned as lower latency/cost than LLMs for decision (not generation) tasks. Powered by a decision model (post mentions TypeSafe AI’s Jev); available in Beta via SQL and REST; compatible with TypeSafe AI API for existing users. “Fraction of a second” / cost claims are vendor. Beta ≠ GA.
What’s New Mon Sep 28, 2026 ~2:29 PM ET (pubDate 18:29 GMT); ML blog dated 28 SEP 2026. AWS adds xAI Grok 4.7 to Bedrock with US Geo and Global cross-Region inference profiles: us.xai.grok-4-7 and global.xai.grok-4-7. 500K token context; reasoning effort low / medium / high / xhigh. APIs: Responses, Chat Completions, Converse / InvokeModel; OpenAI-compatible /openai/v1 path on bedrock-runtime. Emphasize Bedrock availability + profile IDs — do not re-lead Grok 4.7 model launch (Issue 23 / Sep 21). Bench numbers in AWS blog are vendor/third-party.
What’s New Mon Sep 28, 2026 ~11:00 AM ET (commercial + GovCloud). Two access paths — Bedrock (IAM/CloudTrail/Guardrails/Knowledge Bases) and Claude Platform on AWS (Anthropic-native APIs with AWS billing/auth). Global CRIS profile for Bedrock. Companion to Frontier lead; not a separate H1.
Deprecations page: access ends 2026-09-28 for gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106. Recommended replacement listed as gpt-5.6-terra. Next waves: gpt-5.4-cyber → gpt-5.6-cyber by Oct 1, 2026; Oct 23, 2026 legacy chat/reasoning snapshots (gpt-4 / gpt-3.5-turbo-0125 / o1 / o3-mini / o4-mini etc. → gpt-5.6-sol/terra/luna). Grep production configs for the four Sep 28 IDs before traffic fails.
What’s New Fri Sep 25, 2026 ~3:00 AM ET (pubDate 07:00 GMT) — before Issue 25 publish but not carded then. Skills give coding agents validated step-by-step guidance (verify sending identities, build branded RCS agents, send production email). Integrates with Claude Code, Codex, Cursor, Kiro. Install via aws-core plugin or MCP config, then add channel skill. Links setup guides for SMS/RCS, WhatsApp, SES.
Install OpenShell 0.1.x (quickstart pin OPENSHELL_VERSION=v0.1.0; GitHub also has v0.1.2). Run a coding agent (Codex / Claude Code) inside the sandbox; verify FS/process and outbound HTTP/MCP supervisor policy (OPA/Rego). Early OSS — expect rough edges. Quarantine “milliseconds” and BlueField Sentry path are optional / vendor.
Claude Platform release notes September 24, 2026; refusals-and-fallback docs table “as of September 2026.” Newly billed: refusals that arrive before any output when stop_details.category is bio, frontier_llm, or reasoning_extraction. Mid-stream refusals already billed. Charged at rates of the model that ran the request; applies on all platforms (API, Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry). Other pre-output categories (e.g. cyber, general_harms) still not billed; fallback credit unchanged. Refusals still count against rate limits.
AWS What’s New 2026-09-23 (RSS pubDate Wed, 23 Sep 2026 00:06:00 GMT ≈ Tue Sep 22, 8:06 PM ET — after Issue 22, before Issue 23 morning publish, not covered in Issue 23). GA of CloudWatch Omni: OTel + CloudWatch; spaces for multi-account/Region telemetry including Azure workloads; natural-language Q&A over telemetry with AWS DevOps Agent root-cause help; dedicated agent observability / evaluation-driven workflow for LangGraph, CrewAI, OpenAI Agents SDK, Vercel AI SDK, and Strands (prompt / model call / tool invocation eval + experiments). Free IDE extension for VS Code, Cursor, and Kiro (local instrument/debug/eval; no AWS account required per WN). Regions: us-east-1, us-west-2, eu-west-1.
Google Developers Blog 2026-09-23. Antigravity Python SDK supports offline/local agentic workflows. Initial path: Gemma 4 26B A4B via LiteRT (litert-lm import from litert-community/gemma-4-26B-A4B-it-litert-lm; LiteRTAgentConfig). Docs recommend machine with >24 GB VRAM or unified memory. Also LocalOpenAIAgentConfig for OpenAI-compatible servers (Ollama, LM Studio, vLLM). Hybrid demo framing: cloud Gemini 3.8 Flash as planner vs local Gemma swarm; blog claims one recorded run used 95 cloud tokens and 97.2% of tokens (3,322) local — vendor demo numbers, not an independent bench.
September 22, 2026. Genie One MCP server GA. Companion miss (not a re-lead): Unity Catalog Pages (Sep 22) — governed business-concept definitions for Genie Ontology. Optional industry short: Amazon Connect Customer A2A (Sep 22).
