This was the week AI stopped being a tool you open and started being a colleague you deploy. OpenAI shipped its most agentic model yet. ChatGPT and Gemini both turned themselves into enterprise platforms with shared agents, governance hooks, and pre-wired access to your apps and data. Walmart announced AI training and an in-house agentic platform for 2.1 million employees. And Anthropic’s most powerful unreleased model — the one designed to find cybersecurity vulnerabilities — slipped through a third-party vendor. The org chart is being rewritten. Let’s get into it.

On Thursday, April 23, OpenAI launched GPT-5.5, the model it had been quietly calling “Spud.” Greg Brockman framed it as “a big step towards more agentic and intuitive computing.” For once, the numbers behind the marketing line up.

What changed: GPT-5.5 is natively omnimodal — text, image, audio, and video through a single architecture — with a 1M-token context window. OpenAI cites particular advances in four buckets: agentic coding, computer use, knowledge work, and early scientific research. It runs at GPT-5.4 latency for fewer tokens, which is the kind of efficiency gain that quietly reshapes deployment economics.

Where it lives: ChatGPT and Codex from launch day for paid subscribers. API access is gated behind cybersecurity guardrails — OpenAI is no longer pretending the most capable models can ship to the API simultaneously. Pricing: $5 / $30 per million input/output tokens. Double GPT-5.4’s rate, but with computer use and a 1M-token window bundled in.

Why it matters more than the version number suggests: This is the first model where the marketing words “agent” and “computer use” map cleanly onto a feature an enterprise can actually buy and run today. Seven days after Anthropic’s Opus 4.7, GPT-5.5 retook the frontier on coding and agentic benchmarks — but the bigger story is that it shipped as the engine inside ChatGPT’s new Workspace Agents. The model isn’t the product anymore. It’s the substrate.

For enterprise leaders: If your AI roadmap still treats agents as a 2027 question, this week pulled the timeline forward by twelve months. The autonomy bar moved from “demo” to “default” — and the suppliers shipping it are the same ones already inside your Microsoft and Google contracts.

OpenAI quietly turned ChatGPT into something it never quite was before: a place where teams build, share, and govern shared software workers. Workspace Agents launched as a research preview for Business, Enterprise, Edu, and Teachers tiers — free until May 6, then credit-priced.

What you get: Codex-powered shared agents that plug directly into Slack, Salesforce, Google Drive, Microsoft 365, Notion, and Atlassian Rovo. Build once, share with a team, govern through organisation-level permissions. OpenAI’s launch examples — a Software Reviewer that files IT tickets, a Product Feedback Router that monitors Slack, a Weekly Metrics Reporter that pulls data and generates charts — are deliberately mundane. That’s the point.

The succession story: Custom GPTs were a clever consumer feature that never quite became infrastructure. Workspace Agents are the enterprise-grade replacement: they run in the cloud, keep working when you close the tab, and share state across a team rather than a single chat. This is the moment ChatGPT moved from “individual productivity tool” to “team workflow OS.”

The governance reality: Gartner’s first-take note on the launch flagged the obvious: identity, audit, data-loss-prevention, and lifecycle management for agents are now an IT problem, not a vendor demo. Each Workspace Agent is effectively a service account with read/write access to Slack threads, CRM records, and code repositories. Multiply by hundreds of agents across thousands of users and the audit trail gets interesting fast.

For enterprise leaders: Standing up a “head of agentic governance” role is no longer optional. The early movers will write the internal playbook — which agents teams are allowed to share, what data classes they can touch, how they’re decommissioned — before legal writes it for them.

While OpenAI and Anthropic compete on models, Google spent this week doing the unglamorous work of selling AI the way enterprises actually want to buy it. At Cloud Next ’26 in Las Vegas, the message was relentlessly consistent: we sell the agents, the data layer, the runtime, the protocol, and the partner ecosystem.

The three flagship pieces: Gemini Enterprise Agent Platform — build, scale, govern, and optimize agents in a single console. Workspace Intelligence — a context-aware AI layer across Gmail, Docs, Drive, Chat, and Calendar that started rolling out to Rapid and Scheduled domains on April 22. Agentic Data Cloud — a cross-cloud Lakehouse plus Knowledge Catalog that lets agents act on your business data at agentic speed.

The hidden story: Google announced a $750M fund to incentivize its 120,000-member partner ecosystem to build agentic apps, and shipped A2A (agent-to-agent), an open protocol designed to let agents from different vendors actually talk to each other. The protocol is the more important announcement. If A2A gets meaningful adoption, every “AI vendor lock-in” conversation gets a third option: heterogeneous agent meshes that route by capability, not brand.

Why this is the enterprise pitch: CIOs aren’t buying models — they’re buying outcomes wrapped in identity, data residency, audit, and partner support. Google’s pitch this week was: we already have your data (Workspace), your governance (Cloud IAM), your data residency (BigQuery), and now your agent platform. Microsoft’s response will be Build in May. Until then, Google is the only vendor talking like the AI is already a fact of corporate life rather than a future state.

For enterprise leaders: This is the quarter to map your AI stack against the three platform plays — Microsoft + Copilot, Google + Gemini Enterprise, OpenAI + Workspace Agents — and decide where your data already lives. Multi-cloud is fine; multi-platform AI is going to be expensive.

At MIT Technology Review’s EmTech AI Summit late last week, Walmart’s Chief People Officer Donna Morris laid out the largest workforce AI program ever attempted: AI training and certification for all 2.1 million Walmart employees, co-built with OpenAI and Google’s Gemini team, paired with an in-house agentic platform designed for daily use across stores, supply chain, and corporate.

The framing: “People-led, tech-powered.” CEO Doug McMillon told the workforce that “every job will change.” But the explicit, repeated message has been augmentation rather than replacement — agents that find products in stockrooms faster, translate between cashiers and customers, route service queries to the right person, pre-fill paperwork. The agents amplify the associate; the associate stays the operator of record.

The scale story: 2.1 million is more than the entire active-duty US military. Walmart is the largest private employer in the country. When the largest private employer tells its workforce that AI fluency is part of the job, every Fortune 500 HR department recalibrates the floor.

The agentic part: The internal platform isn’t just a chatbot. It’s an agentic layer designed to act inside core retail systems — inventory, scheduling, supplier comms — with associates as the human approvers. Walmart is effectively building the 2.1-million-seat reference architecture for “people working alongside agents.” Every other large enterprise will study this one closely.

For enterprise leaders: The new question on every workforce-strategy slide is no longer “will we train our people on AI?” — it’s “how does our AI training program compare to Walmart’s?” If the answer is “we don’t have one yet,” that’s now the gap to close before the next board cycle.

The week’s cautionary tale came from the AI lab least likely to feature in one. Bloomberg, Fortune, and TechCrunch all reported that a small Discord group of unauthorized users gained continuous access to Mythos — Anthropic’s most powerful unreleased model, designed to find cybersecurity vulnerabilities — and has been using it since shortly after its private launch.

How it happened: At least one member of the group is a third-party contractor for Anthropic. The group reportedly guessed where the model was hosted by combining inside knowledge with deployment patterns leaked earlier from Mercor, an AI training startup. Anthropic confirmed it is “investigating a report claiming unauthorized access to Claude Mythos Preview through one of our third-party vendor environments.”

What Mythos can do: Anthropic publicly demonstrated Mythos finding a 27-year-old security vulnerability in OpenBSD — an operating system specifically renowned for security. The model is designed for enterprise cybersecurity firms and a handful of technology partners; Anthropic has been deliberately cautious about wider release because the same capability that finds vulnerabilities can also weaponize them.

What hasn’t happened (yet): The group has reportedly not used Mythos for cyberattacks. They’ve been using it continuously to probe what it can do. That’s the more interesting story: Mythos can apparently still produce useful output in unauthorised hands without obvious tripwires firing. The internal access pattern looks like the contractor’s normal traffic.

For enterprise leaders: This is the textbook third-party vendor risk story for the agentic era. Your AI vendor’s third-party contractors are now part of your AI vendor’s threat model — and therefore yours. Three things to do this week: (1) audit which AI vendors you depend on for regulated workloads, (2) ask each one for their third-party-vendor controls in writing, (3) add an “AI tooling exposure” line to your vendor risk register, distinct from generic SaaS exposure.

The week’s stories rhyme more than the headlines suggest. GPT-5.5 makes agents a default rather than a feature. Workspace Agents and Gemini Enterprise put those agents directly in the org chart. Walmart trains 2.1 million people to work alongside them. And Mythos showed how quickly the most carefully governed model can leak through one third-party vendor’s misconfigured login.

The pattern from the CIO seat: The question is no longer whether agents are coming — it’s how the existing IT operating model accommodates a new class of “non-human worker” that signs into your apps, reads your data, and makes decisions in your name. The 2025 playbook was “stand up a chatbot.” The 2026 playbook is “run a workforce of agents inside the same identity, audit, and lifecycle systems you run humans through.”

The data behind the tension: Gartner’s most recent enterprise AI adoption work shows agentic AI investments accelerating into 2026 — but governance, audit, and identity controls are trailing materially behind deployment. Forrester’s Q1 2026 buyer-behaviour data tells the same story: enterprises are buying agents faster than they’re standing up the security and lifecycle plumbing to manage them. The Mythos leak is what that gap looks like in production.

The strategic decision that matters most this quarter: Identity. Treat agents as a first-class identity class — not extensions of the human users who created them. That means dedicated service accounts, scoped permissions, audit trails that survive when the human owner leaves the company, and an explicit decommissioning workflow when the agent is retired. Get this right and the rest of the agentic stack composes around it cleanly. Get it wrong and the next “vendor breach” headline could feature your logo.

The forward question: In twelve months, what fraction of the work that flows through your enterprise will originate with an agent rather than a human — and have you decided yet who is accountable for the decisions those agents make? CIOs who answer that out loud this quarter set the agenda. The ones who don’t will inherit it.

That’s your signal for the week of April 20–26, 2026.

AI clocked in this week. If this was useful, subscribe and forward it to one person who’d appreciate it before their next board meeting. See you next week.