Google shipped Gemini 3.5 Flash today at I/O, now live in the Gemini app and AI Mode in Search, with Gemini Omni and the Gemini Spark agent arriving next week for AI Ultra subscribers.

• Gemini 3.5 Flash benchmarks better than Gemini 3.1 Pro on coding and agentic tasks, runs 4x faster than competing frontier models, and is priced at roughly half the cost — Demis Hassabis cited 800 tokens/sec inside the Antigravity framework.
• CodeMender, a security agent that detects vulnerabilities, proposes fixes, validates them, and applies patches with approval, is being integrated into Google Cloud Agent Platform for enterprise use alongside a new AI Content Detection API for synthetic media governance.

Nvidia released Nemotron-Labs-Diffusion today, a language model family that combines autoregressive, parallel diffusion, and self-speculation decoding in a single architecture across 3B, 8B, and 14B sizes.

• The diffusion decoding mode generates 6x more tokens per forward pass than Qwen3-8B, directly attacking the throughput bottleneck that makes autoregressive-only inference expensive at scale.
• The family ships with base, instruct, and vision-language variants, making it deployable across both text and multimodal enterprise workloads without architecture switching.

GitHub expanded the Copilot Cloud Agent with two lower-cost model options — Claude Haiku 4.5 and GPT-5.4 mini — both at a 0.33x cost multiplier, available now per the May 18 changelog.

• The 0.33x multiplier means routine agent tasks like ticket triage, test generation, and documentation now cost roughly one-third of a full frontier model run, enabling organizations to segment workloads by complexity rather than paying frontier rates uniformly.
• The addition of two competing vendors at the economy tier signals GitHub is positioning Copilot as a model-agnostic orchestration layer rather than a single-model product.

Red Hat today announced updates to AI Factory with Nvidia adding hardened security controls for continuously running agents and a production-ready deployment path combining Red Hat AI and Nvidia AI Enterprise.

• The update targets the specific gap between proof-of-concept agent deployments and production: governance, security, and scalable orchestration built on OpenShift with Nvidia's open-source OpenShell project.
• Red Hat is positioning AI Factory as infrastructure for agents that never stop running, which requires different security guarantees than batch or session-based AI workloads.

NHS Shared Business Services went live today with a Salesforce-powered AI platform called SBS One that handles invoice and procurement queries across the NHS through automation, self-service, and AI agents.

• NHS SBS processes shared services for NHS organizations at scale; the deployment targets finance and procurement query handling, the high-volume, low-judgment work that previously required human triage before escalation.
• The stated goal is freeing frontline staff and redirecting capacity to patient care, framing ROI in labor reallocation rather than cost reduction — a politically durable justification for public sector AI investment.