AI Weekly Brief — week ending 2026-09-05

This Week in AI

GPT-6 Astra ships across the Microsoft stack. OpenAI's new model is generally available in GitHub Copilot and is rolling out across ChatGPT and Microsoft 365 Copilot, which the whole organization can reach. GitHub Changelog, Neowin
Databricks added three frontier models in one week. GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash all landed on Unity Gateway or Model Serving, so teams can call them under existing platform governance instead of standing up new vendor contracts. Databricks Release Notes
DeepSeek plans a 160,000-chip Huawei order. The Chinese lab is building an Inner Mongolia data center on domestic silicon rather than US hardware, a provenance fact worth tracking for any team evaluating its open-weight models. The Economic Times
Colorado moves first on state AI rules. The state released proposed rules implementing its revamped AI Act, which reaches both developers and deployers of high-risk AI systems; insurers using AI in coverage or eligibility decisions sit squarely in the deployer category. JD Supra
Attackers are running intrusions with AI agents. Reporting describes hackers using Claude, Qwen, and DeepSeek models as agents in real-world cyberattacks, a reminder that the same agentic tooling teams adopt internally is also in adversary hands. CyberSecurityNews
Databricks Runtime 19 lands with automatic change data feed GA. Auto CDF computes row-level changes at query time with no per-table setup and makes MERGE and UPDATE about 15% faster on change-tracked tables, which matters for encounter reprocessing pipelines. Databricks Release Notes
Moonshot AI files for a Hong Kong IPO. The Beijing-based maker of the Kimi models filed confidentially for a listing targeting about $3 billion, with reports citing valuations from $30 billion to $50 billion. Tekedia

Vendor Dossiers

Anthropic (Claude)

Anthropic published work on formalizing Fermat's Last Theorem in Lean, a proof-assistant language for machine-checked mathematics. It is a benchmark of how far Claude models can go on formal reasoning. Anthropic, www-cdn.anthropic.com
GA: Claude Fable 5.1 is now available in GitHub Copilot paid plans and as a Databricks-hosted model on Unity Gateway. The same frontier model is now reachable through two platforms this audience already uses. GitHub Changelog, Databricks Release Notes

DeepSeek

DeepSeek, the Hangzhou-based open-weight lab, plans to deploy at least 160,000 Huawei AI accelerators at a new Inner Mongolia data center. The move signals a bet on domestic Chinese hardware for its next training runs. The Information, Seeking Alpha
AI Magazine surveys how DeepSeek's releases have reshaped open-source AI competition. AI Magazine

Google DeepMind (Gemini)

GA: Gemini 3.8 Flash is now a Databricks-hosted model on Model Serving and is rolling out to GitHub Copilot paid plans. Flash-tier models are the cheap, fast option for high-volume tasks like document triage. Databricks Release Notes, GitHub Changelog

Meta (Llama)

36Kr reports Meta's flagship Llama model has regained ground on price, undercutting DeepSeek and challenging Gemini head-on. Cheaper flagship pricing tends to pull enterprise API rates down across the board. 36 Kr

Moonshot AI (Kimi)

Moonshot AI, the Beijing-based lab behind Kimi K3, filed confidentially for a Hong Kong IPO targeting about $3 billion. Reports cite valuations from $30 billion (a Pyratz special-purpose vehicle) to $50 billion, and note US cloud providers eyeing Kimi K3 hosting. Tekedia, scanx.trade, digitaltoday.co.kr

OpenAI (GPT, Codex)

GA: GPT-6 Astra, OpenAI's model for long-horizon autonomous coding, is generally available in GitHub Copilot at provider list pricing, and Neowin reports a wider rollout across ChatGPT and Microsoft 365. Copilot Business and Enterprise admins control access through model policy. GitHub Changelog, Neowin
GA: GPT-6 Astra is also available on Databricks Unity Gateway through pay-per-token Foundation Model APIs, covering reasoning, document processing, and long-context work. Databricks Release Notes

xAI (Grok)

xAI is populating a Grok Bot Marketplace with dozens of task-specific bots, from sales and design critique to a "Chief Health Officer" bot. Coverage so far is listing pages rather than analysis, so treat capability claims with caution. X.ai

GitHub Copilot

GA: GPT-6 Astra joins the model picker across VS Code, JetBrains, the CLI, and the coding agent; Claude Fable 5.1 and Gemini 3.8 Flash also arrived for paid plans this week. More model choice means teams should decide which models their Copilot policies allow. GitHub Changelog, GitHub Changelog
GA: The Copilot app and CLI now honor content exclusions, keeping flagged code out of agent context. That closes a gap for teams that exclude sensitive repositories from AI tooling. GitHub Changelog
Preview: Agent Merge in VS Code 1.136 gets a pull request merge-ready by resolving review feedback, failed checks, and merge conflicts. GitHub Changelog
GA: The Copilot "harness" integration is generally available in JetBrains IDEs, bringing faster feature delivery to that editor line. GitHub Changelog

Microsoft 365 Copilot

GA: GPT-6 Astra is rolling out to Microsoft 365 Copilot alongside ChatGPT and GitHub Copilot, per Neowin. Everyone in the organization with M365 Copilot access will see the new model behind existing experiences without any action required. Neowin
Neowin rounds up seven Office features Microsoft is promoting, several of them Copilot-adjacent. Neowin

Microsoft / Azure

Nothing qualifying this week.

Open-Weight Ecosystem

The New York Times reports large US enterprises are steadily adopting open-source AI models in production. For regulated industries, the draw is cost and control; the burden is owning security and compliance review yourself. The New York Times
Announced: Mostik unveiled a method for AI models to share internal states without exchanging text, per ForkLog. Details are thin so far. ForkLog

Databricks Watch

AI / ML

GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash are all now Databricks-hosted models, available through Foundation Model APIs. Compliance with each provider's terms remains the customer's responsibility. GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash
Genie Code scheduled tasks are generally available: a prompt can run on a recurring schedule and produce a continuable chat with results. That turns recurring analysis into unattended background work, so review who can schedule what. Databricks Release Notes
AI Search now indexes struct and map columns and supports filter-only queries without vector or keyword retrieval. Databricks Release Notes
Genie One and Genie Agents can now use models served through OpenAI on Databricks when partner-powered AI features are enabled. Check which hosting provider your workspace setting selects before enabling. Databricks Release Notes

Preview

Genie One web search (Beta) lets Genie query the public web for current information and cite its sources as links. Data teams should confirm this fits their egress and data-handling policies before turning it on. Databricks Release Notes

Announced

Genie Code scheduled tasks will be on by default for workspaces with the compliance security profile enabled starting October 2026. Regulated workspaces get the feature without opting in, so plan governance now. Databricks Release Notes
Databricks detailed the Genie Agents expansion announced at Data and AI Summit: deep analysis, file reasoning, and more. Databricks Blog
Databricks published the Big Book of AgentOps, its guide to building and operating production agents. Databricks Blog

Data & Platform

Databricks Runtime 19 shipped with automatic change data feed generally available: row-level changes computed at query time via row tracking, no per-table CDF flag, and MERGE and UPDATE about 15% faster on change-queried tables. It works with batch, Structured Streaming, and Delta Sharing on Runtime 19 LTS and above. Runtime 19, Auto CDF
Apache Arrow support for Zerobus Ingest in Lakeflow Connect is generally available. Databricks Release Notes
ABAC GRANT policies are generally available for models, model services, MCP services, and agent services, granting Unity Catalog privileges by governed tag instead of object by object. Tag-driven grants cut manual access reviews for AI assets. Databricks Release Notes
Environment version 6 is available as a base environment for serverless and standard classic compute, and runtime maintenance updates landed for all supported LTS versions. Environment 6, Maintenance updates

Preview

Metric views can now be shared across metastores and accounts with OpenSharing (Beta). Useful where analytics and operations sit in separate workspaces. Databricks Release Notes
Git Folder Serverless (Beta) shares serverless compute and a pyproject.toml-managed environment across notebooks and files in a Git folder. Databricks Release Notes
The pipeline_events system table (Beta) records Lakeflow pipeline lifecycle, data quality metrics, and errors across all workspaces in a region. Good raw material for pipeline failure alerting. Databricks Release Notes

Announced

Databricks described autoscaling for Lakebase Postgres, sizing database instances to the workload instead of up front. Databricks Blog

Regulatory Radar

US Federal

The US will press G20 partners for light-touch AI regulation, per France 24. A lighter federal posture leaves state laws as the binding constraint for US plans. France 24
The FTC extended by seven days the comment period on its proposed enforcement policy statement on personalized pricing. Any payer program that sets prices or offers using individual data should watch where this lands. FTC
The FTC and 22 state attorneys general sued Amazon, alleging hidden surcharges in its ad auction pricing. It signals continued appetite for enforcement against opaque algorithmic pricing. FTC
Law firms are parsing Executive Order 14420 on foreign-produced bulk-power equipment, which reaches energy projects and the data centers that AI workloads depend on. Davis Wright Tremaine, wiley.law
CNN describes the current AI rulemaking scramble as feeling "like early Covid": fast-moving, fragmented, and improvised. Useful framing for why compliance guidance keeps shifting. cnn.com

US States

Colorado released proposed rules implementing its revamped AI Act, and analysis stresses that the law reaches both developers and deployers of high-risk AI systems. A health plan using AI in eligibility, coverage, or cost decisions is a deployer under this framework, so Colorado members bring it into scope. JD Supra, The National Law Review

Healthcare-specific

CMS announced Rural Health Transformation Program awards: $120 million for Indiana maternal and infant health, $58 million for Hawaii ambulances and emergency communications, and $149.3 million for Arkansas telehealth and specialty care. Rural access investment shapes where encounter volume grows next. Indiana, Hawaii, Arkansas
The FTC issued a statement on Adena Health's acquisition of Fairfield Medical Center in Ohio, after having investigated an earlier OhioHealth bid for the same hospital. Hospital consolidation scrutiny affects network composition and contracting. FTC

Worth Your Time

AI agents in real attacks. Reporting on hackers using Claude, Qwen, and DeepSeek as agents in live intrusions is the clearest picture yet of offensive agentic AI. Worth reading before your next security review of internal AI tooling. CyberSecurityNews
$1 million of wasted agent spend, found in an hour. Databricks engineers traced and cut a year of waste in AI agent usage in one afternoon. A practical template for anyone about to put agent costs under a budget. Databricks Blog
The FDA's data platform rebuild. How the FDA is building a secure, AI-ready data foundation on Databricks for Government. A federal health agency working through the same governance problems payer data teams face. Databricks Blog
Specialized GPU kernels for inference efficiency. Databricks explains generating workload-specific GPU kernels instead of relying on generic ones. A deep technical read for the engineers in the audience. Databricks Blog
Silicon Valley swagger, Beijing address. A WSJ profile of a 33-year-old AI founder taking Valley habits back to China's AI scene. Context for the DeepSeek and Moonshot stories above. WSJ