Data & AI watch, week of 5 to 11 October 2026

Generally available since July, Planning in Fabric IQ now connects to a semantic model with the signed-in user's identity, with no shared cloud connection to configure; existing permissions apply as they are. Budgets and forecasts are written back into the same semantic model, and event triggers start a pipeline, a Dataflow Gen2, a refresh or a Teams notification on each writeback. Planning sheets can sit in an organisational app alongside Power BI reports. A new calculation engine (preview), which Microsoft says is up to ten times faster, handles tens of millions of cells. Custom fiscal years, new grid layouts and multi-driver simulations round out the release.

The September Fabric feature summary, last updated on 7 October, details several controls for administrators. Generally available: an allowlist of authentication methods for cloud connections (Basic, service principal, OAuth Entra ID), and an allowlist of Entra ID tenants those connections may reach, up to one hundred. Both apply to new and existing connections, from the Power BI admin portal. Also available, a Purview DLP rule can restrict access to a Fabric item as soon as it detects sensitive data in it. In preview, Fabric policies define who can create a given item type or share data externally.

The Fabric Core MCP Server gives AI agents an endpoint hosted by Microsoft, authenticated with Entra ID, with nothing to install locally. An agent can search the OneLake catalog, manage workspaces, items and permissions, and view capacities; every operation respects Fabric roles and is recorded in the audit log. In preview, the Fabric Notebook Toolkit library lets GitHub Copilot CLI, Codex or Claude Code edit, run and diagnose a notebook. On the Git side, branch workspaces are now generally available, with branch out, switch and checkout gathered in the Source control pane.

According to the Microsoft 365 message center (MC1474458), a Power BI report can be added as a reference in a Copilot Notebook, alongside files, to ground a summary, a presentation or a brief in its data. The public preview finishes rolling out in mid-October; general availability starts in late October and completes in mid-November. Nothing to configure: the feature appears on its own wherever Power BI and Copilot are enabled.

Anthropic released Claude Haiku 5.5, available on its API and through AWS, Google Cloud and Microsoft Azure. Pricing depends on request size: up to 100,000 tokens, $0.10 input and $0.50 output per million tokens, and five times that above. The vendor reports an average cost about 75% lower than Haiku 4.5 and much higher scores on its benchmarks, figures that are its own. It is the first Haiku with an adjustable reasoning effort setting, designed for high-volume tasks and as a sub-agent alongside Opus 5.5 and Sonnet 5.5.

OpenAI is rolling out GPT-6 Sol from 7 October on Plus, Pro, Business and Enterprise plans, and GPT-6 Luna from 8 October on Free and Go plans. The model brings Intelligent UI: an answer can combine text, charts, buttons, forms and small calculation tools, depending on the question. OpenAI also says GPT-6 Instant starts answering web-search questions 44% sooner on average, a vendor figure.

Mistral AI opened a preview of Mistral Large 4 on its API: a multimodal model of roughly one trillion parameters, served from its European datacentres, at $1.36 input and $4.18 output per million tokens. The announced performance, notably in cybersecurity, finance and law, comes from the vendor. According to AFP, the model weights will be published on 27 October; the launch follows a €3 billion funding round valuing the company at more than €21 billion.

The European Commission held a special meeting of its scientific panel this week on the safety of frontier AI models, focused on recent loss-of-control incidents; the panel and the AI Office prepared questions for the vendors involved, without naming them. Separately, an overview published on 6 October by ActuIA recalls the timeline set by omnibus regulation 2026/1744: providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to mark the content they produce, and Annex III high-risk requirements apply from 2 December 2027.

On 6 October, Anthropic expanded its Cyber Verification Program: security teams of any size, open-source maintainers and established researchers can apply for access to its most capable models for incident response, malware analysis or vulnerability validation. On 8 October, the Cyber Mission added a free scanner for open-source projects, which sends reports with a proof of concept and a suggested fix, along with funding for the Python Software Foundation, OpenSSF and the Apache Software Foundation.

According to the Financial Times, reported by TechCrunch, OpenAI told investors its annualised revenue is approaching $50 billion, against about $70 billion cited in late September. The two leading vendors calculate this metric differently: Anthropic counts sales made through its cloud partners, OpenAI leaves them out. OpenAI's stock market listing is expected in early 2027.