The week's main developments, grouped by original publication date. Every item links to our full story and the original source.
This week is defined by major enterprise platform updates and ecosystem expansions. Microsoft extended Copilot access with stronger identity governance, automated agent identity, and new billing methods. Google Cloud added OpenAI models to its unified platform and expanded identity and confidentiality options for AI workloads.
Cohere released Rerank 5 in Pro and Fast tiers with Context Focus, which returns relevant passages inside documents and reduces LLM token costs.
Rerank 5 aims to lower retrieval token cost while preserving answer quality, with two speed tiers and broad platform availability.
CALIBER by Cohere reduces a 7B model's calibration error by 52.5% on BigMathDigits.
Reliable confidence estimates help users decide whether to trust a model's output or verify it, which is critical for deploying AI in high-stakes settings.
Microsoft Copilot Cowork is now generally available to all Microsoft 365 Copilot tenants worldwide in tier-1 languages.
Organizations can compare available models, platform governance, and usage-based billing against their own workflows and approval rules.
Cohere's Confidential Computing on Model Vault, with GPU memory protection and Oblivious HTTP encryption, is now available in public beta on Microsoft Azure with NVIDIA H100 GPUs.
Enterprises need verifiable assurance that AI inference remains confidential from cloud and platform operators. Cohere's approach makes the trust boundary explicit and measurable, reducing reliance on implicit infrastructure trust.
Shopify's post-trained model for the Sidekick GraphQL subagent surpasses frontier accuracy and cuts serving costs by 90%.
The numbers come from company-reported benchmarks, not independent measurements; validate similar trade-offs on your own tasks before adopting them.
Brazilian energy firm Vibra cites a 55% productivity rise in prioritized activities and a 23% reduction in manual task hours, equal to 37,920 hours saved per year, using Microsoft 365 Copilot.
This case highlights how leadership involvement and structured data enable AI adoption in large enterprises without compromising security.
Google Cloud now offers a general availability option for Anthropic-managed Claude models, giving teams choice between immediate feature access and standard governance.
This expands choice for developers needing specific data processing or early access to new platform features.
Microsoft details primary model choices for Copilot Studio agents, outlining preview limits, security evaluations, and specific retirement dates for GPT models late 2026.
This helps developers avoid unsupported models and manage data privacy risks when choosing agents based on speed, cost, and safety benchmarks.
Microsoft's new Azure OpenAI Decisions API returns predicate, choice, or score answers from text and images, powered by gpt-6-luna.
The Decisions API enables deterministic classification, routing, and scoring in applications without relying on freeform text generation, reducing complexity and cost for content moderation and triage tasks.
Microsoft Copilot Studio made Claude Sonnet 5 and GPT-5.5 Chat available as primary AI models for agents, improving response quality and reasoning.
These updates mark a shift toward stronger identity governance, new automation capabilities, expanded model choice, and consumption-based pricing in Copilot Studio.
Microsoft Foundry Managed Fine-Tuning gains four open-source models in public preview, introduces MAI-Code-1.1-Flash, and launches Multi-LoRA shared GPU hardware deployments with hourly billing.
These updates reduce hosting costs for multiple fine-tuned variants via shared hardware and expand the range of customizable models, including the first Microsoft coding model for fine-tuning in Azure AI Foundry.
Google Cloud announced October 21, 2026, launching OpenAI GPT-6 models in Model Garden.
Google Cloud integrates OpenAI models into its enterprise platform, giving customers unified access to GPT, Gemini, and Claude with consolidated billing.
Hardware-backed encryption for AI workloads in 36 zones with under 3.5% throughput loss.
Organizations can protect model weights and user data during computation with verifiable hardware isolation, meeting compliance without significant performance trade-offs.
FP8 tensors on MPS: 5–6.5× MoE decoding speedups; Wan2.2 video VAE memory drops from ~270 GiB to 8 GiB; softmax ~8× faster; complex64 LU factorization added.
Release fixes performance barriers for on-device AI on Macs. Dedicated Metal kernels make large language and video models more practical.