OpenAI opened GPT-5.6 Sol (flagship), Terra (mid-tier), and Luna (a new budget tier at $1/$6 per million tokens) to every ChatGPT, API, and Codex user on July 9, ending a 13-day government-coordinated preview that began June 26. SpaceXAI publicly launched Grok 4.5 the same day, a 1.5-trillion-parameter V9 foundation model with supplemental training data from Cursor, priced aggressively at $2/$6 per million tokens, 60 to 80% under Sol. Anthropic separately restored and extended Claude Fable 5 access through July 12 for Pro, Max, Team, and select Enterprise plans, while Claude Sonnet 5 is now the default model worldwide with a 1 million token context window at promotional $2/$10 pricing through August 31.

Claude also reached general availability on Microsoft Azure AI Foundry this week, the first Claude deployment on NVIDIA GB300 Blackwell Ultra GPUs, and Anthropic launched a Claude Marketplace with ten launch partners including Snowflake, letting existing spend commitments carry across platforms.

What this means for enterprise leadership: model selection is no longer an annual decision, it is now a live, recurring one. Three credible, materially different options landed on the same day this week alone. Organizations without a documented, repeatable evaluation and switching process are accumulating governance exposure every time a new model drops, whether they choose to switch or not.

McKinsey's 2026 AI Trust Maturity Survey, roughly 500 organizations with direct AI governance, risk, or investment responsibility, fielded December 2025 through January 2026, finds only about one-third reach maturity level 3 or higher in strategy, governance, and agentic AI controls. The average Responsible AI maturity score rose to 2.3 in 2026 from 2.0 in 2025, genuine progress, still short of maturity.

For enterprise leadership: this is not a call to slow spending, it is a call to close the gap between spend and oversight before it widens further. This week's model-launch story (Section 01) makes the abstract governance number concrete: every organization evaluating GPT-5.6, Grok 4.5, or an open-weight alternative this week just generated a new governance decision, whether or not anyone documented it.

Agentic AI in banking moved decisively from pilot to production infrastructure in 2026, not just conversation. Fiserv launched agentOS; FIS introduced a Financial Crimes AI Agent built with Anthropic. Banks are concentrating deployments on procedural, auditable work, financial-crime detection, regulatory-change triage, controls testing, and continuous transaction monitoring, in governed environments where every agent decision is traceable and outputs require human approval.

In healthcare, 61% of organizations are already building, implementing, or have secured budget for agentic AI, and 85% plan to increase investment over the next two to three years. Yet only 3% report agents live in production. Mount Sinai Health System and Mayo Clinic are both using agentic tools to streamline workflows and automate repetitive tasks toward more personalized care, and the UK's NHS launched a system-wide project focused specifically on responsible, collaborative, and sustainable agentic AI deployment, a public-health-system governance template worth watching.

For financial services and healthcare leadership: in both sectors, budget and executive buy-in are no longer the binding constraint. The gap is a documented, auditable path from pilot to production, exactly what regulators in both sectors are now signaling they expect to see.

Hyperscaler AI capital expenditure was raised to $750 billion for 2026, up from $670 billion, on pace to cross $1 trillion in 2027. New AI-ready data center capacity keeps stacking up: Equinix's DC01UK near London won local approval with a £4 billion ($5.3 billion) build commitment, Pure DC secured $2.7 billion for AI infrastructure across Europe and the Middle East, and Galaxy Data Center raised $250 million for a Southeast Asia platform.

Meta is using autonomous construction robots to help build the solar farm powering its own Hyperion AI data center, a striking illustration of the convergence. NVIDIA and LG are unifying model development, physical AI data generation, robot simulation and training, edge deployment, and factory-scale digital twins into a single workflow, and NVIDIA's Jensen Huang and Hyundai's Chung Euisun agreed to deepen their alliance across mobility, manufacturing, and robotics through Hyundai's Boston Dynamics subsidiary.

For manufacturing and energy leadership: energy, robotics, and agentic AI are no longer three separate planning tracks. A robot building the solar farm that powers the data center running the agents is the clearest single example yet that these belong in the same capital-planning conversation.

Colorado now joins California in anchoring a disclosure-and-rights US-state model, distinct from the EU's risk-tiering approach. Any organization operating across multiple US states should confirm which model its governance documentation was built around, because the three approaches now require materially different evidence.

Every Thursday, Field Notes Live puts one live, timely question to our network. This week's question follows directly from Sections 01 and 02: three frontier labs just gave every enterprise a live model-selection decision to make, and most organizations still lack a documented governance process for making it.