Data & AI watch, week of 26 October to 1 November 2026

Version 3000.338 of the on-premises data gateway ships Power Query engine 2.159.229: a report published and then refreshed through the gateway runs with the same logic as the October release of Power BI Desktop. A connection test now keeps your Power Query settings intact, and a failed service start comes with detailed diagnostics. In preview, new admin REST APIs give Fabric administrators, or a service principal, visibility over every connection in the tenant, with each action recorded in the audit log. Microsoft reminds you that it supports the last six releases of the gateway.

The Govern tab of the OneLake catalog is now generally available. It brings together domains, capacities, workspaces, tags and policies, with health indicators and recommended actions: endorsements, sensitivity label coverage, inactive items. The underlying report, in the admin monitoring workspace, is built on a semantic model you can extend, whatever your licence or tenant size. Govern Skills, published in Skills for Fabric, let you run these audits with an agent such as GitHub Copilot. In preview since 29 October, a policy defines which workspace admins can change customer-managed key encryption and outbound networking settings.

Born from the acquisition of Osmos in January 2026, Project Osmos takes a complete problem and the expected outcome, then works through investigation, PySpark code, real execution in a Fabric Spark session and validation, reporting its evidence as it goes. Before it starts, you set the guardrails: allowed resources, output destination, staging tables, schema changes. The agent acts with your permissions, consumes Fabric capacity units and is launched from GitHub Copilot CLI, Claude Code or Codex; trial capacities are excluded. On 26 October, the Fabric Notebook Toolkit (FNTK) opened the same path for notebooks: a coding agent runs one cell, reads its output and corrects it, with no need to resubmit the whole notebook.

With workspace monitoring, completed queries from warehouses, lakehouse SQL analytics endpoints and warehouse snapshots land in the WarehouseExecutions table of the monitoring KQL database: duration, status, result code, CPU time. You can compare warehouses with each other and spot slow or failed queries from one place. On 28 October, T-SQL scalar functions gained BREAK and CONTINUE in loops, plus a validation step that explains why a function cannot be inlined into the query.

Fabric Real-Time Dashboards gain, as generally available features, layout tiles (title, text box, labelled divider), multi-KPI groups computed from a single query, and KPI styling that shows status, scale and target.

On 26 October, Anthropic filed its S-1 with the SEC to list its Class A shares on Nasdaq under the ticker ANT, with Goldman Sachs and Morgan Stanley. The number of shares and the price range are yet to be set. The filing shows revenue of $16.3 billion in the first half of 2026, against $4.6 billion for the whole of 2025, and more than 6,000 customers spending over $100,000 a year. The half-year net loss of $80.7 billion comes almost entirely from financial charges; the operating loss is $693 million. An amendment was filed on 2 November.

OpenAI is rolling out GPT-6.1 Terra in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans, and in the API as gpt-6.1-terra. Its standard price is half that of GPT-5.6 Terra. According to OpenAI, it matches GPT-6 Astra on a test of agents using workplace software, at a tenth of the price per token, and finishes tasks more than three times faster on average than Astra and GPT-6.1 Sol. These performance figures come from the vendor.

Google presents Gemini 4 Carbon as its most advanced model, especially for cybersecurity and long-running tasks. It opens this week to a subset of cyber defenders, then to trusted partners through the Fairwind programme; developers and businesses will follow, with no date announced. Launch pricing: $2 input and $10 output per million tokens, then $4 and $20. Google reports 88.2% on SEC-bench Pro for vulnerability finding, measured without guardrails; these are vendor figures.

Announced for the end of October, the open weights of Mistral Large 4 have been available on Hugging Face since 29 October: around one trillion parameters, about fifty billion of them active per token, text and image input, and a one-million-token context. The licence, derived from MIT, allows commercial use and customisation; only reselling the model as a service requires a commercial licence. Through the API, pricing is $0.70 input and $2.10 output per million tokens. Mistral states that the model was trained entirely on European infrastructure; its benchmark results are vendor figures.

Since 26 October, new projects on the paid tier of the Gemini API store generateContent, streamGenerateContent and batchGenerateContent requests by default (store=true). Existing projects keep their setting. You can opt out per request with store=false, or for the whole project in the Logs and Datasets dashboard. On 30 October, the Computer Use tool, which lets the model operate a screen, became generally available.

Anthropic has opened a Usage and Cost Analytics API in beta for Claude Console and Claude Enterprise organisations, filterable by model, workspace, API key and user or service account. The previous reporting APIs remain supported. On 30 October, Anthropic announced the end of fast mode for Claude Opus 4.8 on 30 November 2026; fast mode remains available on Claude Opus 5.5.