Microsoft released Decision-1, a small model fine-tuned from Alibaba's open-weight Qwen3.5-9B that selects between fixed answer options rather than generating free text, priced at $0.042 per million input tokens.

• Built for classification tasks: Decision-1 replaces generative completions with a constrained-choice architecture, which Microsoft says beats rival decision models on accuracy and latency in its own benchmarks.
• Migration path planned: Microsoft says it will port the post-training approach onto MAI and OpenAI models, signaling Decision-1 is a technique template as much as a standalone product.

GitLab 19.4 ships a public-beta /goal command in Duo CLI that accepts an open-ended objective and keeps working until the task completes or hits an iteration limit, without requiring predefined flows.

• Three open-weight models hosted: GitLab now serves its own hosted open-weight models alongside expanded MCP tool coverage, reducing dependency on external API providers for agent execution.
• Spending controls added: New per-project AI cost limits let teams cap token spend before it hits a budget line, addressing a recurring complaint as agentic sessions grow longer.

Slack shipped shareable skills, letting any Slackbot skill travel outside the originating org via a link so external users can add it to their own Slackbot without rebuilding it.

• Marketplace implication: Shareable links create a de-facto distribution layer for Slackbot automation, analogous to app-store sharing but without a formal review gate or revenue split.
• Announced October 8, live now: The feature was announced at the start of the eligibility window and is in production across all Slack tiers.

Google DeepMind's AlphaProof Nexus cracked nine of 353 open Erdős problems and proved 44 OEIS conjectures, with results published in Science after peer review.

• Two problems unsolved for 56 years: The system resolved problems from the Erdős set that had been open since 1970, and DeepMind shared proofs publicly before the journal paper dropped.
• Peer review matters here: Unlike most AI math claims, these results cleared standard academic scrutiny, setting a higher evidentiary bar for future benchmark comparisons.

New York consulting firm JB&B used CrowdStrike Falcon Shield to lift one SaaS application's security posture score from roughly 30% to nearly 100% and save more than 10 hours per week on ongoing review.

• Zero to ten-plus apps covered: Before Falcon Shield, JB&B had no centralized visibility into SaaS security risk; the deployment expanded coverage to more than 10 applications.
• One-time onboarding gain: The team also recovered more than 100 hours of manual effort during the initial assessment and onboarding phase, a concrete ROI anchor for similar-sized professional services firms.

Thompson Rivers University reduced daily security threat review from most of the workday to a few morning hours after consolidating onto Microsoft 365 A5 with Sentinel.

• Small team, open campus: TRU operates an inherently porous research and teaching network with a lean security staff, making consolidated threat visibility a prerequisite rather than an upgrade.
• AI readiness as co-goal: The M365 A5 deployment also established the governance foundation TRU needed before expanding its AI program, linking security posture to AI adoption readiness.