AI Intelligence
What changed, and why it matters.
A curated radar across six AI ecosystems. Every update links to its primary source and carries a short note on what it means for architecture, cost, governance or productivity.
Latest briefing
Copilot Autopilot, reusable skills and Projects
Reusable, shareable skills turn individual prompting into organisational process assets. Someone needs to own curation, versioning and retirement of those skills.
Mistral Large 4 released as open weights
A credible European open-weight frontier model strengthens sovereign AI options for EU enterprises with data-residency and lineage requirements.
Gemini 4 Carbon announced with phased access
Staged, trusted-partner releases are becoming the norm for frontier models. Plan for capability gaps between what is announced and what your organisation can actually procure.
All updates
16 updates
- 30 Oct 2026Microsoft CopilotAgents● High impact
Copilot Autopilot, reusable skills and Projects
Microsoft detailed October Copilot updates: a redesigned Home combining Chat and Cowork, Projects for shared context, reusable skills, and Autopilot, an always-on cloud agent under approval safeguards.
Source: Windows ReportEnterprise impactReusable, shareable skills turn individual prompting into organisational process assets. Someone needs to own curation, versioning and retirement of those skills.
- 30 Oct 2026OpenAI / ChatGPTModels
OpenAI Introduces Safety Report and Teen Protections
OpenAI is enhancing ChatGPT with expanded safeguards and earlier interventions to prevent violence. The company has also introduced a new commitment to public safety reporting.
Source: OpenAI NewsEnterprise impactThis change may improve trust in AI systems for enterprise use, particularly in sectors requiring high levels of user safety and compliance.
- 29 Oct 2026OpenAI / ChatGPTResearch & Policy
Updated Model Spec organised around five values
OpenAI reorganised its Model Spec around faithfulness to human intent, harm minimisation, honesty and transparency, human agency and care, adding guidance on context, oversight and agency.
Source: OpenAIEnterprise impactA useful reference when writing internal AI usage policies: it frames how vendor models are meant to behave when instructions are silent, which matters for agentic deployments.
- 29 Oct 2026Microsoft CopilotEnterprise
More than 25 finance skills across Copilot
Microsoft expanded Copilot for finance with more than 25 new skills and partner plugins for accounting and capital markets, grounded in organisational context through Work IQ.
Source: Microsoft Copilot BlogEnterprise impactDomain-specific skill packs are arriving for core functions. Finance leaders should pilot against close and reporting processes with explicit review steps.
- 29 Oct 2026Open ModelsModels● High impact
Mistral Large 4 released as open weights
Mistral released the weights of Mistral Large 4, a 1-trillion-parameter multimodal model trained on European infrastructure, leading open-weight coding benchmarks.
Source: Mistral AIEnterprise impactA credible European open-weight frontier model strengthens sovereign AI options for EU enterprises with data-residency and lineage requirements.
- 29 Oct 2026Open ModelsModels
Granite Switch 4.1.1 Enables Model Updates Similar to Software
The Granite Switch 4.1.1 update allows users to update models as easily as software, improving maintainability and scalability. This change streamlines the process of applying updates without requiring full retraining or system overhauls.
Source: Hugging Face BlogEnterprise impactThis update simplifies model maintenance in enterprise environments by enabling seamless upgrades, reducing downtime and operational complexity.
- 28 Oct 2026Google / GeminiModels● High impact
Gemini 4 Carbon announced with phased access
Google DeepMind announced Gemini 4 Carbon, its strongest model with leading cybersecurity and long-horizon capabilities, initially available to a subset of cyber defenders.
Source: GoogleEnterprise impactStaged, trusted-partner releases are becoming the norm for frontier models. Plan for capability gaps between what is announced and what your organisation can actually procure.
- 27 Oct 2026OpenAI / ChatGPTEnterprise
Oracle applications come to ChatGPT
OpenAI and Oracle are bringing Oracle applications and data into ChatGPT, with Oracle developing Oracle AI Workspace for approved business workflows governed by existing permissions.
Source: OpenAIEnterprise impactThe assistant is becoming the front door to ERP. Expect pressure to define which finance, procurement and planning workflows may be executed from a chat surface, and under whose approval.
- 26 Oct 2026OpenAI / ChatGPTModels● High impact
GPT-6.1 Terra: fast computer use at half the price
OpenAI introduced GPT-6.1 Terra, a fast, efficient model combining visual reasoning, precise computer use and clearer writing. Standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
Source: OpenAIEnterprise impactHigh-volume back-office automation (invoice capture, form filling, drawing interpretation) becomes economically viable. Re-run the business case for any workflow previously rejected on token cost.
- 21 Oct 2026Anthropic / ClaudeAgents
Claude Duet: a proactive agent in research preview
Claude Duet is a research-preview app where Claude works as a proactive agent across connected apps, sending a morning brief and checking before acting on the user's behalf.
Source: Claude by AnthropicEnterprise impactProactive agents shift the governance question from "what can a user ask" to "what may an agent do unprompted". Approval checkpoints become a design requirement, not an option.
- 20 Oct 2026PerplexityAgents● High impact
Perplexity Computer runs fully local on Mac
Perplexity Computer can now run model, harness and task execution entirely on Apple silicon Macs, using Qwen 3.6 or PPLX Pro, with cloud access only when authorised.
Source: PerplexityEnterprise impactLocal execution offers a practical answer for sensitive data and cost control. It strengthens the case for hybrid architectures where private work stays on device.
- 20 Oct 2026PerplexityEnterprise
Computer on sovereign, air-gapped infrastructure
Perplexity partnered with Valarian to run Computer inside institutions' own servers, including air-gapped networks, for regulated industries.
Source: PerplexityEnterprise impactRegulated sectors gain an agentic option that respects sovereignty constraints. Evaluate alongside private model hosting rather than as a SaaS decision.
- 15 Oct 2026Open ModelsModels
Meta releases Muse Spark Open weights
Meta released Muse Spark Open, an 816B-parameter (42B active) multimodal model with a 1M-token context window, available through major clouds and inference providers.
Source: Meta AI ResearchEnterprise impactLarge open-weight models are now available on every major cloud. The decision shifts from "open or closed" to licence terms, hosting cost and support model.
- 8 Oct 2026Anthropic / ClaudeEnterprise
Claude Dashboards connect to enterprise data platforms
Claude Dashboards (beta) builds live dashboards from platforms such as BigQuery, Databricks and Snowflake, showing the query behind every chart.
Source: Claude by AnthropicEnterprise impactSelf-service analytics moves into the assistant. Data governance, semantic definitions and access controls on the warehouse become the real quality gate.
- 8 Oct 2026Google / GeminiEnterprise● High impact
Gemini agent: one universal agent for work
At Gemini at Work 2026 Google announced the Gemini agent inline across Gmail, Drive, Docs, Sheets and Calendar, with agent identity, permission controls, smart routing and spend caps.
Source: Google CloudEnterprise impactGoogle is positioning a single agent with shared memory across the suite. Spend caps and agent identity are now table stakes to demand from any vendor.
- 7 Oct 2026Anthropic / ClaudeModels● High impact
Claude Haiku 5.5: around 75% cheaper than Haiku 4.5
Anthropic released Claude Haiku 5.5 for high-volume, cost-sensitive work such as summaries, classification and live support, and halved Sonnet 5.5 cache-read prices.
Source: AnthropicEnterprise impactTiered model portfolios (small model as sub-agent, frontier model as planner) now have a clear cost case. Revisit routing rules in existing agent designs.
Intelligence tracks the change. Essays explain what it means.