AI Weekly Brief

AI Weekly Brief — week ending 2026-08-18

This Week in AI

  • DeepSeek raised prices, sharply. DeepSeek released V4 Pro at 14x the price of its cheapest model and moved to peak-hour API pricing, a shift away from the cheap-token strategy that made it famous. The era of loss-leader API pricing may be ending (Forbes, Table.Briefings).
  • Alibaba's Qwen went small and passed 3 billion downloads. Alibaba released Qwen 3.8 27B, a model sized to run on laptops, days after its download count passed Meta and Google in open models. Small local models keep getting more capable, which matters for any future on-premises or edge use case (CNBC, Silicon Republic).
  • Microsoft is reshuffling Copilot. The consumer and enterprise Copilot apps are merging into one app while Microsoft retires several features, including Excel's COPILOT function and the free Deep Research tool. Expect UI and feature churn in tools your teams use daily (Windows Central, The Register).
  • Copilot security flaws drew headlines. Researchers showed a one-click path to exfiltrate data from apps connected to Microsoft Copilot Personal, and a separate report detailed a hidden input that let Copilot be manipulated. A reminder that AI assistants wired into email and files widen the attack surface (The Hacker News, Ars Technica).
  • Model routing became the cost lever of the week. Databricks announced Smart Routing in Unity AI Gateway, claiming frontier quality at 30%+ lower cost per task, and Snowflake shipped dynamic model routing with the same pitch. Platforms are now competing on picking the cheapest adequate model for you (Databricks, TechTarget).
  • Databricks moved a wave of sharing features to GA. SecureConnect (sharing data behind firewalls), managed Iceberg table sharing, and sharing tables on default storage all went generally available, with default enablement rolling out through early 2027. Details and billing changes below in Databricks Watch (Databricks release notes).
  • OpenAI published a cluster of security and safety posts. It described new safeguards that pace frontier model development against cyber-capability risk, a defender-focused security outlook, and a teen-specific ChatGPT with parental controls (OpenAI, OpenAI, OpenAI).

Vendor Dossiers

Alibaba (Qwen)

  • Released: Alibaba launched Qwen 3.8 27B, a model sized to run natively on laptops, positioned directly against Meta's open models and targeting edge deployment (CNBC, AI Business).
  • Qwen models passed 3 billion downloads, moving ahead of Meta and Google in open-model distribution. Download counts are a rough proxy, but the ecosystem shift is real (Silicon Republic, Dealroom).
  • Announced: Apple reportedly trained a China-specific model for Apple Intelligence with Alibaba's help, and is bringing Qwen to Mac users in China. A marker of how AI stacks are splitting along geographic lines (gHacks, MarketScale).

Anthropic (Claude)

  • Anthropic published work on Claude autonomously designing de novo protein binders, part of a broader push into protein design and analytical chemistry. Relevant as a signal of where frontier labs see scientific-workflow value, including in life sciences (Anthropic, Anthropic).

DeepSeek

  • Released: DeepSeek launched V4 Pro at 14x the price of its cheapest model and raised API prices overall, including peak-hour pricing. The pivot from loss-leader pricing to commercialization changes the cost math for anyone benchmarking against DeepSeek's APIs (Forbes, Global Times).
  • Released: DeepSeek open-sourced DeepSeek Harness, an MIT-licensed agent harness in developer preview where every component is a plugin. Another permissively licensed building block entering the coding-agent tool chain (Open Source For You, MarkTechPost).

Google DeepMind (Gemini, Gemma)

  • Announced: Google partnered Gemini and Pixel with five global football clubs for AI-driven matchday features. A consumer marketing play, not a platform change (Google).

Kakao (Kanana)

  • Announced: Kakao's Kanana-2 topped peer models in a safety evaluation (thelec.net).

OpenAI (GPT, Codex)

  • Announced: ChatGPT for Teens launched with stronger built-in protections, healthy-use features, and parental controls. Notable as a template for how vendors segment users under safety pressure (OpenAI).
  • OpenAI detailed new monitoring, alignment, and security safeguards that pace frontier model development against cyber-capability risk, alongside a companion post on what AI means for attackers and defenders. Useful context for security teams planning around AI-assisted threats (OpenAI, OpenAI).
  • Announced: OpenAI launched an initiative to support democratic oversight of AI in national security, offering tools and training to government institutions (OpenAI).
  • Asana used Codex to replace an outdated testing system in two weeks for about $12K, work it had scoped at five years. The claim is vendor-published, but it is the kind of legacy-modernization case data engineering teams should watch (OpenAI).

xAI (Grok)

  • Announced: xAI published documentation for Grok Bot aimed at teams and enterprises (xAI Docs.
  • Trade coverage put Grok's capability gains alongside Gemini's speed improvements in a widening performance-and-price race (thelec.net).

GitHub Copilot

  • GA: Enterprise managed settings arrived in Copilot for JetBrains, covering plugin governance, MCP server allowlists and denylists, OpenTelemetry configuration, and permission modes. This closes a governance gap for organizations standardizing Copilot controls across IDEs (GitHub Changelog).

Microsoft 365 Copilot

  • Announced: Microsoft began merging the consumer and enterprise Copilot apps into one unified app, and removed the Windows 11 "Ask Copilot" taskbar feature as part of the pivot. Expect the Copilot entry points your teams use to shift (Windows Central, CyberSecurityNews).
  • Announced: Microsoft is retiring several Copilot features, including the COPILOT function in Excel, with users told to back up affected files. If anyone on your team built spreadsheet workflows on the COPILOT function, they need a migration plan (MakeUseOf, The Register).
  • Announced: Copilot's free Deep Research capability is ending, with its successor available only on paid tiers. Budget owners should note the pattern of free capabilities moving behind paywalls (Notebookcheck).
  • WARNING: Security researchers disclosed flaws in Copilot Personal that could let a single click exfiltrate data from connected apps, and a separate report showed a hidden input used to manipulate Copilot. Worth flagging to security teams reviewing Copilot connector policies (The Hacker News, Ars Technica).

Microsoft / Azure

No significant developments this week.

Open-Weight Ecosystem

  • Hugging Face published a walkthrough of multi-vector (late interaction) embedding models in Sentence Transformers, relevant to anyone building retrieval over documents (HuggingFace).
  • IBM Research examined how much memory an agent actually needs, a practical read for teams sizing agent architectures (HuggingFace).
  • A GPU-management case study showed a 33-point utilization gain on the same cluster from reordering workloads, with lessons for anyone paying for shared compute (HuggingFace).

Databricks Watch

AI / ML

GA

  • Google Gemini 3.7 Flash is now available as a Databricks-hosted model through Foundation Model APIs (release notes).
  • Genie One: workspace-level chat instructions are GA, and the interface now accepts image and Word uploads, PDF attachments for scheduled tasks, and MCP writes (Beta) (release notes).
  • Genie Code now supports file uploads as chat context (release notes).

Preview

  • Genie Agents picked up several Beta features: content search on attached volumes, volume file analysis without Unity AI Gateway, prompts alongside uploaded files, and API improvements (release notes).

Announced

  • Databricks announced Document Intelligence for extracting structured data from messy documents. Directly relevant to shops with paper-adjacent intake like faxes, PDFs, and scanned forms (Databricks blog).
  • Smart Routing in Unity AI Gateway routes each task to the cheapest model that meets quality, claiming 30%+ lower cost per task at frontier quality (Databricks blog).

Data & Platform

GA

  • OpenSharing SecureConnect is GA: providers can share data from storage behind firewalls or private endpoints without per-recipient network allowlists. Note the expanded networking billing that comes with it, including new charges billed to recipients (release notes, billing details).
  • Sharing managed Iceberg tables is GA, including sharing to external Iceberg clients (release notes).
  • Sharing tables backed by default storage is GA and being enabled by default on a rolling basis through early 2027 (release notes).
  • The SQL alert task in Lakeflow Jobs now supports all compliance security profile standards and is on by default for workspaces with the profile enabled. Useful for regulated workspaces that gate job logic on data-quality alerts (release notes).

Preview

  • Identity attributes (Beta): nine IdP-sourced attributes such as department and cost center can now drive Unity Catalog ABAC column-mask policies. A finer-grained option for masking sensitive columns by user attributes rather than group membership alone (identity attributes, ABAC policies).
  • Session restore for serverless jobs (Beta): failed, long, or cancelled runs snapshot their Python and Spark state so you can debug in a notebook without rerunning the job (release notes).
  • Base environments now work on classic compute (Beta) for managing Python dependencies (release notes).
  • Lakeflow Connect added three Beta connectors: Workiva, SendGrid, and Gmail (Workiva, SendGrid, Gmail).
  • The Tags page in Governance Hub (Beta) gives account admins a central view of tag usage, invalid values, and untagged assets (release notes).

Announced

  • The root_cause_analysis and downstream_impact columns in the data quality monitoring results system table are being deprecated, starting August 18 and September 7 respectively. Check any queries or dashboards that read these columns (root_cause_analysis, downstream_impact).
  • Azure Databricks Runtime 10.4 LTS reaches end of life on November 1, 2026. Workloads still pinned to it need to move to a supported runtime (Azure Updates).

Regulatory Radar

US Federal

  • The FTC won a court decision blocking Henkel's acquisition of Liquid Nails, keeping the Loctite maker from buying its main construction-adhesive rival. No direct AI or health-plan impact; it signals continued active merger enforcement (FTC).
  • Bill-payment firm Doxo will pay $2.1 million over FTC allegations it used search ads to impersonate billers and hid add-on fees. Relevant to any payer whose members reach payment or portal pages through search: impersonation and fee-disclosure practices are under enforcement scrutiny (FTC).
  • The FTC sent $23.8 million in refunds to drivers and diners harmed by Grubhub's deceptive earnings claims and fee practices. Another data point that consumer-facing claims and fee disclosures carry real enforcement cost (FTC).

Worth Your Time

  • Bloomberg — Why China's affordable AI worries Silicon Valley, and what this week's DeepSeek price hike does to that story.
  • Just Security — "Test, Standardize, Restrict": a proposed US policy framework for Chinese AI models, useful background if your vendors start bundling them.
  • TechTarget — Snowflake's dynamic model routing, the clearest sign yet that AI cost control is becoming a platform feature rather than a spreadsheet exercise.
  • Databricks blog — How the Databricks Feature Store serves features with sub-second freshness; a solid engineering read for anyone running fraud or eligibility models on fast-moving data.
  • Databricks blog — A plain-language explainer on AI hallucinations, worth sharing with analysts who are new to reviewing model output.