Anthropic announcement Anthropic priced Opus 5 at $5 per million input tokens and $25 per million output tokens, and said that is the same price as Opus 4.8. Anthropic announcement Anthropic also introduced a Fast mode, saying it runs around 2.5 times the default speed and is priced at twice Opus 5’s base price. Anthropic announcement Enterprises that still treat model choice as an annual procurement event will lag. Your workflows are going to want dynamic routing across models based on risk tier, latency, and cost per task. Anthropic announcement The operational question for your CTO and COO is whether you have an evaluation harness that can tell you, every week, whether the “cheaper frontier” is actually cheaper after rework, hallucination containment, and review time. If you do not measure that, you will optimize tokens and miss outcomes.

Model Context Protocol Blog update The project says the revision introduces a “stateless core” and removes the initialize handshake and protocol-level session, enabling servers to scale behind a simple round-robin load balancer. Model Context Protocol Blog update A separate briefing on the revision says the final spec ships July 28, 2026 and that sessions, the initialize handshake, and the Mcp-Session-Id header are removed. David & Goliath briefing The strategic implication is that the center of gravity moves from bespoke agent hosting to standardized gateway control. Once the transport is boring, the value shifts to policy, identity, tracing, and audit. That is where your differentiated governance posture will live. If your team is building MCP servers today, the governance question becomes: can you prove which tool was invoked, by which agent, under which identity, for which user intent. Stateless scaling makes that easier, but it also increases the volume of actions. Your controls have to keep up.

Microsoft Source press release Microsoft said Mistral Medium 3.5 and OCR 4 are now available in Microsoft Foundry, and that Mistral Medium 3.5 is now in Copilot Studio. Microsoft Source press release Microsoft said organizations can deploy Mistral models across cloud, cloud-connected, and fully disconnected environments, and it described a new multibillion-dollar agreement focused on expanding AI infrastructure in Europe. Microsoft Source press release The strategic question for enterprises is whether you are building a portable deployment pattern, or a vendor-specific one. “Disconnected” deployment is only durable if your evaluation harness, your policy routing, and your incident response also work in that mode. For Microsoft and Mistral, the deeper move is market structure. They are packaging model choice, security, and deployment posture as one integrated offering, which pulls governance buyers into the platform layer instead of letting them shop the model layer independently.

ServiceNow Q2 2026 financial results ServiceNow said Anthropic became the first design partner, connecting Claude directly to ServiceNow workflows and actions. ServiceNow Q2 2026 financial results ServiceNow also said AI Control Tower expanded with new discovery, observation, governance, security, and measurement capabilities to give enterprises control over every AI system, agent, and workflow regardless of where it runs. ServiceNow Q2 2026 financial results If you are an enterprise buyer, the lesson is that your system of record is becoming your system of action arbitration. The “integration” problem becomes a policy problem: which agents can trigger which actions, with which evidence, and with which human signoff. If you are a vertical SaaS vendor, this is the playbook you have to answer. Either you become an action gateway with real accountability assets, or you get enveloped into a callable component behind somebody else’s assistant.

Ted Lieu and Nathaniel Moran introduced the bipartisan AI Kill Switch Act, which would require covered AI developers to maintain controls that can throttle inference, restrict users or capabilities, suspend service, or shut down a system. Shelly Palmer Palmer wrote that the draft applies to developers with at least $500 million in relevant annual revenue and AI systems whose training compute would cost over $100 million, and that penalties for noncompliance could reach $20 million per day. Shelly Palmer For enterprise operators, the immediate action is continuity planning. If your business logic depends on one model endpoint, a government-ordered throttle becomes a direct operational outage. The backup plan cannot be “call the vendor.” It has to be an alternate routing path, an alternate model, and an alternate operating mode. The strategic pattern: the governance buyer is turning into the platform buyer. The vendor who can supply audited controls, throttles, and portable fallback will win regulated enterprises, even if their baseline model is not the absolute best.