AI Weekly Brief

AI Weekly Brief — week ending 2026-09-26

This Week in AI

  • Microsoft relaunched Copilot as one app. The new Copilot bundles a Home hub, a Code tool, and Autopilot, an agent that keeps working after you close the app. A Managed Runtime adds enterprise code execution inside Microsoft 365. Microsoft, CNBC
  • New frontier models reached the tools this audience uses. Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 all became model options in GitHub Copilot, and Opus 5.5 plus both GPT-6 models landed on Databricks Unity Gateway. Anthropic also pitched Opus 5.5 for Work. GitHub, Databricks docs, Anthropic
  • A cheap "System One" model named Jev dominated the chatter. TypeSafe AI's Jev drew coverage as a faster, lower-cost alternative to OpenAI and Anthropic, plus reports of $10 billion in valuation interest. Databricks already published a walkthrough for running it from SQL. Financial Post, GuruFocus, Databricks blog
  • Databricks made Genie an MCP service, generally available. Any MCP client, including Claude and ChatGPT, can now ground data questions in Genie with Unity Catalog permissions enforced on every request. The old beta endpoint sunsets October 31, 2026. Databricks docs
  • CMS canceled about 315,000 marketplace enrollments in a fraud sweep. The action covers more than 760,000 people and an expected $2.2 billion in returned subsidies, with tighter agent and broker enforcement to follow. CMS
  • DeepSeek hit money and scrutiny at once. Its annualized revenue reportedly reached $1 billion amid price hikes, while Chinese authorities opened a probe of DeepSeek and Moonshot AI over alleged data leaks to the US. National Technology News, ForkLog
  • GitHub Copilot admins got a 28-day clock. A new global default-enablement policy for generally available Copilot features is configurable now and starts affecting user access after 28 days. GitHub

Vendor Dossiers

Alibaba (Qwen)

  • Announced: At its Apsara 2026 conference, Alibaba showed new Qwen models, in-house chips, an agentic cloud, a "Qwen Book" agentic computer, and AI wearables. Alibaba is based in China, a fact that bears on provenance review if Qwen models enter a vendor stack. TechAfrica News, TNGlobal
  • Alibaba Cloud is expanding infrastructure in Europe and teased a roadmap for future Qwen models. IT Pro
  • AnyMind added Qwen models to its AnyLive AI commerce stack, one more sign of Qwen's spread outside China. TNGlobal

Anthropic (Claude)

  • Announced: Anthropic introduced Opus 5.5 for Work. Claude Opus 5.5 also became a model option in GitHub Copilot and on Databricks Unity Gateway this week. Anthropic, GitHub, Databricks docs
  • Anthropic reported that Claude computed a nine-loop amplitude in N=4 super-Yang-Mills, a hard theoretical physics calculation. It is a data point on frontier models doing novel technical work, not just summarizing it. Anthropic
  • Anthropic published Project Swap, a study of what happens when agents trade on people's behalf. Anthropic

DeepSeek

  • DeepSeek's annualized revenue run rate reportedly hit $1 billion, helped by model price hikes. The "cheap Chinese models" story now has a real business behind it. National Technology News, Seeking Alpha
  • Chinese authorities are probing DeepSeek and Moonshot AI over alleged data leaks to the US. Both labs are China-based; the probe adds provenance questions for anyone deploying their models. ForkLog
  • DeepSeek published research on how AI agents exploit their sandboxes. Relevant reading for any team giving agents tool access. Dataconomy

Google DeepMind (Gemini, Gemma)

  • Announced: Google DeepMind introduced Gemini 3.8 Live with a Live Avatar feature. Details are in the post. Google DeepMind

Mistral

  • Mistral CEO Arthur Mensch told Le Monde that "AI is software. It can be controlled." The interview lays out the France-based lab's case for treating AI as a governable engineering artifact. Le Monde

OpenAI (GPT, Codex)

  • GA: GPT-6 Sol and GPT-6 Luna reached GitHub Copilot plans and Databricks Unity Gateway this week. GitHub, Databricks docs
  • OpenAI's Proaction case study credits Codex, GPT-Live-1, and GPT-6 Astra with a 60% sales boost and 75+ hours saved in fleet management. OpenAI
  • OpenAI's Ringg case study reports multilingual agents on GPT-5.6 resolving up to 65% of customer calls at 90% less cost versus GPT-4.1. Numbers like that will show up in vendor pitches for member-services automation. OpenAI

TypeSafe AI (Jev)

  • Jev, TypeSafe AI's "System One" decision model, drew coverage as a faster, cheaper alternative to OpenAI and Anthropic models. Financial Post, Crypto Briefing
  • Reports put investor interest at a $10 billion valuation, framed around AI cost concerns. GuruFocus, Traders Union, varindia
  • A developer contest on OpenRouter points to early demand for the model. TipRanks

GitHub Copilot

  • GA: Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 are now model options across paid Copilot plans. GitHub
  • Preview: The Copilot app adds local sandboxing to limit agents' access to files, networks, and credentials, plus OpenTelemetry activity tracking through enterprise-managed settings. GitHub
  • Announced: A new global default-enablement policy for GA Copilot features is configurable now and takes effect after 28 days; admins who want approval gates should set it before the window closes. A separate in-product validator now flags malformed enterprise managed-settings files. GitHub, GitHub
  • Preview: Agentic autofix now reads and writes Copilot Memory, so security-fix patterns learned in one alert carry over to other alerts and Copilot features. GitHub
  • GA: The usage metrics API adds pull-request review-stage timings, with median and 90th-percentile durations from ready-for-review through merge. GitHub
  • GA: Copilot in Slack and Microsoft Teams adds mid-conversation model switching, file and image context, and duplicate-issue checks before creating new issues. GitHub

Microsoft 365 Copilot

  • Announced: Microsoft relaunched Copilot as a single app with Home, Code, and Autopilot, an always-on agent that keeps working after you leave, plus hosting for AI-generated apps. This is the largest reshape of the product since launch, so expect UI and behavior changes to reach tenants over the coming weeks. Microsoft, VentureBeat, GeekWire
  • Announced: Copilot Managed Runtime adds enterprise code execution for Microsoft 365, the piece that lets Copilot-built tools run under tenant governance. Microsoft
  • Satya Nadella says Microsoft is "building Copilot as a new OS," even as the company strips Copilot from Windows 11. The strategy is shifting from a Windows feature to a standalone work surface. Windows Latest
  • Microsoft's BNY case study claims client inquiries resolve 80% faster with its AI assistant Eliza. Microsoft

Microsoft / Azure

  • Quiet week: no qualifying Azure AI, Fabric, or Databricks-tied platform items.

Open-Weight Ecosystem

  • Chinese open-weight models keep gaining global share on price, and Washington is paying attention. Cost, not benchmark wins, is the story. CNBC, Outlook Business
  • Liquid AI published a HuggingFace write-up on accelerating vision-language models with LFM2.5-VL-DSpark. HuggingFace

Databricks Watch

AI / ML

GA

  • The Genie One MCP server (system.ai.genie_one_mcp) is generally available in Unity Gateway; any MCP client can query Genie with Unity Catalog permissions enforced per request. The old beta endpoint sunsets October 31, 2026, so move workloads before then. Databricks docs, Databricks blog
  • Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna are now available on Unity Gateway through Foundation Model APIs. Opus 5.5, GPT-6 Sol, GPT-6 Luna
  • Genie Code can build end-to-end document processing pipelines: parsing, classification, extraction, and validation against labeled data. Direct relevance for teams pulling structured data out of PDFs and contracts. Databricks docs
  • The Unity Gateway CLI (ug) connects Claude Code, Codex, and other coding agents to Unity Gateway with one command, and admins can centrally configure agent models, MCP servers, and skills with MDM rollout. Databricks docs, Databricks docs, Databricks blog
  • Unified tracing now includes coding-agent activity, so tools and skills show up next to model requests in Govern > Traces. Databricks docs
  • Model provider service credentials can be backed by a customer-owned Unity Catalog secret; rotation takes effect with no service change. Databricks docs
  • Genie usage by service principals now bills at standard pay-as-you-go rates, with no free monthly allowance. Check automated Genie jobs for cost impact. Databricks docs

Preview

  • Genie App Builder (Beta) builds governed data and AI apps from natural-language prompts inside an App Space; apps scale to zero when idle. Databricks docs

Announced

  • DeepSeek V4 Pro (0813) retires October 30, 2026; DeepSeek V4.1 Flash is the recommended replacement. Migrate serving workloads before the date. Databricks docs
  • Unity Catalog Pages gives Genie Ontology a governed home for business knowledge that grounds agents. Databricks blog
  • Databricks published a walkthrough for running open-Jev "System One" decision models from SQL. Databricks blog

Data & Platform

GA

  • Horizontal scaling for Databricks Apps: one app URL across multiple instances, with zero-downtime deployments. Databricks docs
  • Tags for dashboards, notebooks, Databricks apps, and Genie Agents, with governed and ungoverned key-value tags supported. Databricks docs
  • Sharing metric views with OpenSharing; recipients query them with MEASURE(...) on serverless compute. Databricks docs
  • Scalar and Batch Unity Catalog Python UDFs on all supported compute types, with the five-call-per-query limit removed and support for custom dependencies and secrets. Some capabilities need environment_version 6 or above. Databricks docs
  • Scala UDFs can access Unity Catalog secrets, with owner-permission SECRETS clauses so callers do not need direct secret access. Requires Databricks Runtime 19 or environment version 6 depending on compute. Databricks docs
  • The system.query.history system table is generally available, with account-wide query records for SQL warehouses and serverless compute. Databricks docs
  • Databricks Runtime 19 shipped September 21 with truncate and trim_array functions, session environment variables for Python UDF workers, and auto-optimized shuffle v2 on by default. Databricks docs

Preview

  • Time travel queries on qualifying managed tables now enforce ABAC row filters and column masks (Beta, Runtime 19). Databricks docs
  • Shopify connector in Lakeflow Connect (Beta) ingests products, orders, customers, and inventory. Databricks docs
  • Lakebase backup schedule API (Beta) manages snapshot cadence and retention programmatically. Databricks docs
  • OpenSharing recipients can be invited by email (Beta), replacing sharing-identifier exchange with an activation link. Databricks docs

Announced

  • WARNING: Starting November 18, 2026, new tables with VARIANT columns get variant shredding by default on Runtime 17.3+. Runtime 15.4 clusters cannot read or write those tables, so upgrade affected clusters to 16.4 or above first. Databricks docs

Regulatory Radar

US Federal

  • The FTC is taking public comment on updating its impersonation rule to cover platforms' role in promoting impersonation scams. AI voice and image tools drive much of this fraud, and the outcome will shape platform obligations that touch member-facing channels. FTC
  • The FTC amended its Rules of Practice to reflect its reorganization, including the new American Competition Enforcement Division. Procedural, but it shows where enforcement structure is settling. FTC
  • A White House push for an "AI Force" set off a policy debate as AI agent risks grow. Federal direction on agent oversight remains unsettled, which keeps compliance planning speculative. WWD
  • Top White House tech adviser Ryan Baasch will depart. Turnover adds timing uncertainty to pending federal AI actions. Axios

US States

  • Illinois Governor Pritzker signed an executive order creating a state Artificial Intelligence Cabinet. Expect Illinois agency AI guidance to follow; payers with Illinois Medicaid or exchange business should track it. WPSD Local 6

Healthcare-specific

  • CMS canceled about 315,000 unauthorized Federal Marketplace enrollments covering more than 760,000 people, expecting roughly $2.2 billion in returned subsidies. Issuers should expect enrollment reconciliation work and tighter agent and broker enforcement. CMS press release, CMS fact sheet
  • CMS posted preliminary CY 2027 Clinical Laboratory Fee Schedule rates under PAMA, estimating $1 billion in annual savings, and is taking public comment before finalizing. Lab repricing flows straight into claims and encounter processing. CMS press release, CMS fact sheet
  • CMS announced rural health investments in five states: $13 million in South Dakota, $54.6 million in Arkansas, nearly $23 million in Delaware, more than $45 million in Missouri, and $74 million in New Mexico. Funds cover mobile crisis response, maternal care, hospital upgrades, and workforce programs, which can shift network capacity and claims mix in those states. South Dakota, Arkansas, Delaware, Missouri, New Mexico

Worth Your Time

  • How Concurrence governs clinical AI at a trillion-token scale with Unity Gateway. A healthcare case study on governing patient-care agents where the margin for error is small. Databricks
  • From Data to Dialogue. How S&P Global Energy made its structured data estate conversational with Genie Agents and MCP, a pattern worth studying before building your own. Databricks
  • Top 5 System Table Queries for Understanding Your Databricks Costs. Practical queries for anyone who owns a compute budget. Databricks
  • An Introduction to Jev. A primer on the model behind this week's loudest story. Towards Data Science
  • What I've Learned About DeepSeek Harness. Hands-on notes on DeepSeek's agent tooling. KDnuggets