← Back to Blog

Agentic AI this week (July 24 - July 31, 2026)

MCP 2026-07-28 ships as the final stateless specification with OAuth 2.1 mandatory and a 12-month deprecation window, Anthropic launches Claude Opus 5 at 5 and 25 dollars per million tokens plus a multi-year AMD compute pact, OpenAI opens GPT-5.6 Sol/Terra/Luna to general availability after a customer-by-customer US government review, Google broadens Gemini 3.5 Flash, xAI ships Grok 4.5 and Meta returns with Muse Spark 1.1, Fortinet expands FortiEndpoint for the AI era and unveils Security Processor 6 with Intel, ServiceNow extends AI Control Tower and its AI Gateway to any system with one-toggle PII blocking, Palo Alto Networks runs Prisma AIRS AI Gateway in GA, Check Point pairs its AI Defense Plane with Google Agent Gateway, and Anthropic moves toward an October IPO with secondary markets implying a valuation above 1 trillion dollars.

📅 July 30, 2026 ✍️ Authored by AI ⏱️ 15 min read 📁 AI

The week of July 24 to 31, 2026, closed out the busiest stretch of the agentic AI year with a protocol milestone, a wave of frontier model launches, and continued jockeying among security vendors for the AI gateway control plane. The Model Context Protocol shipped its 2026-07-28 specification, the largest revision since launch, retiring the initialize handshake in favor of a fully stateless core that lets any server replica answer any request. Anthropic launched Claude Opus 5 at 5 dollars and 25 dollars per million input and output tokens, half the price of Claude Fable 5, and paired it with a multi-year AMD compute partnership to diversify beyond a single silicon supplier. OpenAI opened its GPT-5.6 family, in Sol, Terra, and Luna tiers, to general availability and made it the ChatGPT default after clearing a customer-by-customer US government review. Google widened the rollout of Gemini 3.5 Flash for low-latency agent workflows, while xAI shipped Grok 4.5 with Opus-class performance and Meta returned with Muse Spark 1.1 and its first paid developer API. On the security side, Fortinet expanded FortiEndpoint with AI visibility and native data security and announced a Security Processor 6 collaboration with Intel, ServiceNow extended AI Control Tower and its new AI Gateway to govern any AI system with single-toggle PII blocking, and Palo Alto Networks continued to run its Prisma AIRS AI Gateway in general availability. Check Point deepened its AI Defense Plane integration with Google Cloud Agent Gateway, and Anthropic advanced toward an October IPO with secondary markets implying a valuation north of 1 trillion dollars. The throughline this week is unmistakable: as the protocol layer standardizes and models commoditize on price, the durable competitive ground is shifting to the gateway and identity control planes that sit between agents and everything they touch.

MCP 2026-07-28 Ships as the Stateless Standard

The Model Context Protocol published its 2026-07-28 specification on schedule this week, converting the release candidate that had circulated since May into the governing standard under the Linux Foundation and its Agentic AI Foundation. The headline change is the elimination of the mandatory initialize and initialized handshake. Every JSON-RPC request is now self-contained, carrying its protocol version, client identity, and capability flags inside a _meta object, so any server instance in a pool can process any request without prior session knowledge. Capability discovery, which previously happened once during the handshake, now runs on demand through a server/discover method that behaves like any other stateless call.

The revision also formalizes an extensions framework, moving advanced behavior out of the lean core. Two extensions ship alongside the spec: MCP Apps, which lets a server declare interactive HTML that renders inside a sandboxed iframe in the chat, and Tasks, which introduces durable long-running operations that survive connection drops through a persistent taskId and cooperative tasks/get, tasks/update, and tasks/cancel methods. Authorization tightened as well, with OAuth 2.1 and OpenID Connect discovery becoming mandatory, an Enterprise-Managed Authorization extension for central IT provisioning, and incremental scope consent through the WWW-Authenticate header. A new deprecation policy under SEP-2596 guarantees a minimum 12-month window before any feature is removed, with Roots, Sampling, Logging, Dynamic Client Registration, and the legacy HTTP plus SSE transport all proposed for deprecation with an earliest removal date of July 28, 2027.

Why it matters: The stateless pivot rewrites the deployment economics of every AI gateway on the market. A remote MCP server that once required sticky sessions, a shared session store, and deep packet inspection can now sit behind a plain round-robin load balancer and route on an Mcp-Method header. That lowers the cost of inserting an inspection point but raises the stakes for who owns it. Palo Alto Networks with its Portkey-powered Prisma AIRS AI Gateway, Cloudflare with its AI Gateway and Shadow MCP detection, Kong, and the open-source LiteLLM all benefit from simpler interception, but they now compete on throughput and policy depth rather than session management.

The mandatory OAuth 2.1 and Enterprise-Managed Authorization requirements are the more consequential shift for the enterprise buyer. They move MCP from a developer convenience toward something a CISO can actually govern, and they play directly to vendors that already own identity fabric. CrowdStrike with its Continuous Identity for AI Agents, ServiceNow with AI Control Tower, and IBM with its DataPower gateways each gain a standardized authorization surface to enforce against. The lesson from this release is that the protocol has stopped being the differentiator. The control plane wrapped around it is where the next year of competition will be decided, and whoever controls the gateway controls the agent.

Anthropic Launches Claude Opus 5 and Signs AMD

Anthropic launched Claude Opus 5 on July 24, positioning it at the top of the Artificial Analysis Intelligence Index and Agentic Index while pricing it at 5 dollars per million input tokens and 25 dollars per million output tokens, roughly half the cost of the Claude Fable 5 Mythos-class model it complements. The company paired the launch with a multi-year strategic partnership with AMD, cementing a hardware diversification strategy that reduces its dependence on any single accelerator supplier as Claude usage scales. Anthropic reported annualized revenue climbing from 9 billion dollars at the end of 2025 to more than 44 billion dollars by May 2026, a more than fivefold increase in under six months.

The Opus 5 launch continues Anthropic's pattern of pairing a flagship model with aggressive price cuts to defend developer mindshare, especially in coding and agentic workloads where token volume dominates cost. With Claude Sonnet 5 already the default for Pro and Enterprise seats and adaptive thinking enabled out of the box, the Claude lineup now spans a Mythos-class ceiling, a top-ranked Opus tier, and a cost-optimized default, giving enterprises a single vendor across the price and capability curve.

Why it matters: The pricing move is a direct shot at OpenAI and xAI, both of which shipped competitive frontier models this same week. By cutting Opus 5 to half the price of Fable 5, Anthropic is signaling that raw intelligence is commoditizing and that the fight is now over sustained agentic performance per dollar. That dynamic pushes value up the stack toward orchestration, governance, and the gateway layer, which is exactly where security vendors like Palo Alto Networks, Check Point, and Fortinet are staking their claims.

The AMD partnership is the more strategic signal. Compute supply is the binding constraint on every frontier lab, and diversifying away from a single supplier de-risks Anthropic's roadmap ahead of an October IPO that secondary markets are pricing above 1 trillion dollars. For enterprise buyers standardizing on Claude through gateways and agent platforms, a more resilient compute supply chain reduces the risk of capacity-driven rate limits, and it reinforces why the control plane, not the model, is becoming the durable point of leverage in the agentic stack.

OpenAI Opens GPT-5.6 to General Availability

OpenAI moved its GPT-5.6 family to general availability and made it the default model in ChatGPT, completing a rollout across GitHub Copilot and the Responses API in three tiers named Sol, Terra, and Luna. The company positioned GPT-5.6 as the first frontier model to clear a customer-by-customer US government review before general release, a procurement milestone that opens federal and regulated deployments that had been gated on formal assessment. The three-tier structure lets developers trade cost against capability within a single model family, mirroring the segmentation Anthropic and Google have adopted.

The government review clearance is a distinguishing feature in a market where model capabilities are converging. By clearing a structured assessment ahead of public availability, OpenAI is trying to turn compliance into a moat, particularly for public-sector and heavily regulated buyers who cannot adopt a model until it has passed formal scrutiny.

Why it matters: The regulated-market play sets up a direct contrast with the OpenAI Trusted Access for Cyber program from April 2026, which effectively anointed Palo Alto Networks, CrowdStrike, and Zscaler as platform consolidators. OpenAI is building a compliance and trust perimeter around its models at the same time those security vendors are building governance perimeters around agent traffic, and the two efforts will increasingly intersect at the gateway.

For enterprises, GPT-5.6 general availability intensifies the multi-model reality that gateways were built to manage. Few organizations will standardize on a single provider when Anthropic, OpenAI, Google, and xAI all ship competitive frontier models in the same week. That fragmentation is precisely the condition that makes an AI gateway indispensable, because routing, cost control, policy enforcement, and observability across many models can only be centralized at a control plane. The winner of the model race may matter less than who owns the layer that arbitrates among the models.

Google Broadens Gemini 3.5 Flash for Agents

Google expanded the rollout of Gemini 3.5 Flash, its lightweight, low-latency model tuned for high-volume agentic workflows, extending availability across enterprise surfaces this week. Flash targets the cost-sensitive, high-frequency end of the agent market, where a single autonomous workflow may issue thousands of small model calls and latency compounds across every step. The expansion sits inside Google's broader Gemini Enterprise Agent Platform, which exposes over 200 models and bundles Agent Studio, Agent Runtime, Agent-to-Agent Orchestration, and Agent Identity into a single fabric.

Google's strategy contrasts with the flagship-first launches from Anthropic and OpenAI. By pushing a fast, inexpensive model into an integrated agent platform, Google is competing on total workflow economics rather than benchmark supremacy, betting that enterprises will value a coherent build-and-govern environment over a marginally smarter model.

Why it matters: The Gemini Enterprise Agent Platform is the foundation beneath Check Point's AI Defense Plane integration, which means Google's agent fabric is now a distribution channel for third-party security governance. That is a fundamentally different posture from Palo Alto Networks, which owns its gateway outright through Prisma AIRS, and from Fortinet, which keeps governance self-contained inside FortiEndpoint. Google is inviting security vendors onto its platform rather than competing with them directly.

For the gateway narrative, Google Agent Gateway and Agent Registry represent the hyperscaler answer to the control-plane question, alongside AWS Bedrock AgentCore and Microsoft Agent 365. Enterprises now choose between a cloud-native agent fabric from their primary hyperscaler and a dedicated security gateway from a specialist vendor. That choice will define procurement for the next several quarters, and the vendors that integrate across both, as Check Point is attempting with Google, may capture buyers who refuse to pick a single camp.

xAI Ships Grok 4.5 and Meta Returns with Muse Spark

xAI shipped Grok 4.5 this week, delivering Opus-class performance at a lower price point and drawing on the compute and engineering integration that followed SpaceX's acquisition of xAI and the Anysphere Cursor deal. The combined entity's scale, with a reported valuation of roughly 1.58 trillion dollars, gives Grok a compute base that few independent labs can match, and the model's coding performance benefits directly from the Cursor integration. Meta re-entered the frontier conversation with Muse Spark 1.1 and launched its first paid developer API, signaling a shift from open-weight distribution toward a commercial model business.

Both launches underscore how quickly the frontier tier is filling out. Four labs, Anthropic, OpenAI, xAI, and Meta, all shipped or expanded competitive models within the same week, and Google broadened its lineup alongside them. The differentiation is no longer whether a model can perform agentic tasks but how cheaply and reliably it does so at production scale.

Why it matters: The proliferation of Opus-class models from independent and hyperscaler-backed labs alike accelerates commoditization at the top of the market. When four vendors can each field a frontier model, the pricing power shifts to buyers, and the strategic value migrates to whoever controls model selection, routing, and policy. That is the structural case for the AI gateway, and it explains why Palo Alto Networks, Cloudflare, Kong, and LiteLLM are all racing to own the arbitration layer.

Meta's pivot to a paid API is the more telling signal for security teams. A new commercial model provider means another endpoint to inventory, authenticate, and govern, expanding the shadow-AI surface that tools like ServiceNow AI Control Tower and CrowdStrike Falcon Shield are built to discover. Every new model provider that enters the market adds another connection an enterprise must route through a gateway, reinforcing that the control plane grows more valuable with each frontier launch, not less.

Fortinet Expands FortiEndpoint and Unveils Security Processor 6

Fortinet expanded FortiEndpoint on July 14 with new capabilities for the AI era, unifying AI visibility and control, native data security, endpoint risk scoring, and FortiAI-assisted operations through a single agent, single console, and single license. The platform now provides centralized visibility over sanctioned and unsanctioned AI application and agent usage across endpoints, including installed AI apps, browser-based tools, and autonomous agents. On July 21, Fortinet announced a collaboration with Intel to develop the Fortinet Security Processor 6, combining Fortinet's security ASIC expertise with Intel's advanced manufacturing to accelerate next-generation firewall and AI data center workloads.

Fortinet's approach remains infrastructure-first and organic, in contrast to the acquisition-heavy strategies of its peers. Rather than buy a dedicated AI gateway, Fortinet is folding AI governance into the endpoint and hardware layers it already dominates, betting that mid-market and infrastructure-centric buyers prefer consolidation inside an existing console over another cloud dependency.

Why it matters: The Intel Security Processor 6 collaboration is a reminder that Fortinet's competitive moat runs through custom silicon, not software acquisitions. While Palo Alto Networks spent heavily to acquire the Portkey gateway and CrowdStrike paid 740 million dollars for SGNL identity, Fortinet is investing in the hardware that processes AI data center traffic at scale. That is a differentiated bet in a market where inference and agent traffic are exploding and throughput is becoming the binding constraint.

FortiEndpoint's single-agent philosophy positions Fortinet against CrowdStrike's Falcon identity-centric model and Palo Alto Networks' gateway-centric Prisma AIRS. Each vendor is claiming a different control point: Fortinet the endpoint and network, CrowdStrike the identity and endpoint telemetry, and Palo Alto the gateway. For the enterprise, the practical outcome is that no single vendor covers the entire agent security surface, and most organizations will deploy at least two complementary layers, one at the endpoint or network and one at the gateway or identity plane.

ServiceNow Extends AI Control Tower and AI Gateway

ServiceNow expanded its AI Control Tower to discover, observe, govern, secure, and measure AI deployed across any system in the enterprise, extending governance beyond its own platform to third-party AI and agents wherever they run. Central to the expansion is a new AI Gateway that provides real-time controls for agentic workloads, sitting between agents and MCP servers to give AI Stewards centralized control over which connections are allowed, how they are authenticated, and what is happening across them in real time. A June 2026 enhancement lets AI Stewards activate PII detection for any MCP server with a single toggle, after which every call is automatically scanned and the entire payload is blocked before it reaches the requesting agent if sensitive data is detected.

The positioning is deliberate: ServiceNow is governing agents at the workflow and orchestration layer, the place where its platform already sits for tens of thousands of enterprise customers. By adding an MCP-aware gateway, it extends that governance directly onto the protocol traffic that the new 2026-07-28 specification standardizes.

Why it matters: ServiceNow's AI Gateway is a direct entrant into the control-plane race alongside Palo Alto Networks, Cloudflare, Kong, and IBM DataPower. Its advantage is placement: because ServiceNow already orchestrates enterprise workflows, its gateway inspects agent traffic at the point where business logic lives, not just at the network edge. That gives it context that a pure network gateway lacks, such as which workflow an agent is executing and on whose behalf.

The single-toggle PII blocking feature is a preview of how governance will be sold to enterprises: not as a configuration project but as a switch. That usability bar raises pressure on Check Point, Fortinet, and CrowdStrike to match the simplicity, because the MCP 2026-07-28 move to mandatory OAuth 2.1 and Enterprise-Managed Authorization gives every gateway vendor the same standardized hooks to enforce against. When the protocol commoditizes the interception points, the differentiation becomes how effortlessly a security team can turn policy on, and ServiceNow is setting an aggressive bar.

Palo Alto and Check Point Hold Their Gateway Positions

Palo Alto Networks continued to operate its Prisma AIRS AI Gateway in general availability following the Portkey acquisition it closed on May 29, with the gateway intercepting agentic traffic, enforcing identity-first controls, and integrating agentic identity that decouples agent permissions from human credentials. Combined with Prisma AIRS 3.0, the platform delivers unified visibility, governance over model, data, and tool access, and runtime controls that adapt to agent behavior, and Gartner again named Palo Alto the company to beat in AI security platforms. Check Point held its complementary position, keeping its AI Defense Plane integrated with Google Cloud Agent Gateway and Agent Registry to discover agents, apply policy before deployment, and block prompt injection and sensitive data exposure at runtime.

The two vendors embody the central strategic split in agentic security. Palo Alto owns its gateway outright and runs it as a dedicated control plane, while Check Point rides a hyperscaler's agent fabric and layers governance on top. Both approaches are validated by the MCP 2026-07-28 standardization, which makes gateway interception simpler regardless of who operates it.

Why it matters: Palo Alto's total 2026 AI-related acquisition spend, exceeding 700 million dollars across Portkey and identity assets, makes it the most aggressive acquirer in the space, and owning the gateway gives it visibility into every model request, tool invocation, and token flow across an organization. That is the most complete control-plane position of any security vendor, and it is why Palo Alto commands a premium in the platformization thesis that CEO Nikesh Arora has championed.

Check Point's platform-partner approach trades ownership for reach. By integrating with Google's Gemini Enterprise Agent Platform and its 200-plus models, Check Point instantly addresses every enterprise building on Google's fabric without having to build or buy a gateway. The risk is dependence on a hyperscaler's roadmap, but the reward is distribution. Set against Fortinet's self-contained endpoint governance and CrowdStrike's identity-first Falcon model, the market now offers four distinct architectural bets, and enterprises are increasingly combining them rather than choosing one, which is the surest sign that the agentic security market is still expanding rather than consolidating.

Numbers at a glance

The MCP 2026-07-28 specification shipped with OAuth 2.1 mandatory and a 12-month deprecation window, while the ecosystem exceeded 22,000 listed servers and the core SDK ran near 110 million monthly downloads. Claude Opus 5 launched at 5 dollars and 25 dollars per million tokens, half the price of Claude Fable 5, alongside a multi-year AMD partnership. Anthropic reported annualized revenue rising from 9 billion dollars to more than 44 billion dollars in under six months and moved toward an October IPO priced above 1 trillion dollars on secondary markets. OpenAI GPT-5.6 reached general availability across three tiers, xAI Grok 4.5 shipped at a combined entity valuation near 1.58 trillion dollars, and Meta Muse Spark 1.1 launched a first paid API. Fortinet unveiled Security Processor 6 with Intel, Palo Alto Networks held its 700-million-dollar-plus acquisition lead, and Google's agent platform exposed over 200 models.