OpenAI shipped the equivalent of a full product roadmap in 72 hours.

Same week, Anthropic's supposedly untouchable secret model got cracked open by a Discord group that guessed the URL.

And the U.S. government officially put China on blast for the largest coordinated AI theft campaign in history.

Normal Monday stuff.

Whether building with AI, buying it, or just trying to keep up — this week shifted the context for all of it. Here's what matters.

OpenAI launched GPT-5.5, GPT Image 2, Workspace Agents, and a stack of Codex upgrades. Three days. One company.

GPT-5.5 is live in ChatGPT and the API. Same speed as GPT-5.4, but smarter — scoring roughly 85% on GDPval, which measures AI agents against human experts across 44 real occupations.

Then there's GPT Image 2. Sheesh.

Free tier included. Two-K resolution, near-production typography, multilingual text, up to eight images from one prompt. And the text actually renders correctly now. That changes everything about what it's useful for.

Workspace Agents landed inside ChatGPT for Team plans only — Business, Enterprise, EDU, Teachers. Cloud-running, deployable into Slack, no device required. Free until May 6, then pricing unknown.

Codex got Chronicle (opt-in screen memory, research preview), browser control, Sheets and Slides skills, and OS-wide dictation.

What it means: ChatGPT stopped being just a chat tool this week. Cloud agents running across teams without a device on is infrastructure, not a feature. GPT Image 2 alone opens up use cases that weren't viable a week ago — presentations, mockups, infographics where the text actually reads right. And May 6 is coming fast.

The White House Office of Science and Technology Policy issued a memo accusing China of running deliberate, industrial-scale campaigns to steal U.S. AI models. First reported by the Financial Times.

OSTP Director Michael Kratsios says Chinese entities are using tens of thousands of proxy accounts and jailbreaking techniques to extract outputs from U.S. frontier models — then training cheaper versions on those outputs.

That's distillation. It's doing your AI homework using someone else's answers.

OpenAI, Google, and Anthropic all flagged this publicly in February. Now the federal government is confirming it in writing and pledging to share intelligence directly with U.S. AI companies.

No outright ban on Chinese models yet. But this memo is the loudest the warning has ever been.

What it means: Enterprise teams eyeing Chinese AI models for cost savings now have a federal memo sitting in that conversation. Not a ban — but real institutional weight. The specific risk isn't just IP theft. It's that distilled models often have safety and security protocols stripped out in the process. That's the detail that should land hardest for anyone running AI in a regulated or security-conscious environment.

Google announced Workspace Intelligence at Google Cloud Next — and it's the Gemini feature that's been missing for two years.

Gemini can now access emails, chats, docs, sheets, slides, and Drive across all of Workspace. Context-aware drafts that mirror past preferences and company templates. Working in a doc, it pulls style choices and past edits automatically.

Gmail got AI Inbox too — a task-centered view replacing the standard inbox — plus AI Overviews, which surfaces short summaries of email threads. AI Inbox was previously locked to Google's highest-priced plan.

No firm timeline. Rolling out to Workspace customers gradually.

What it means: This is the version of Gemini people actually wanted since Bard became Gemini. One AI that carries context across the whole workspace — emails, docs, slides, sheets — instead of starting over every time. AI Inbox moving off the ultra-premium tier is the quiet move to watch. For teams on Gmail all day, that's the one that changes daily workflow fastest.

Yuuuuup. Google is going all in.

Ten billion now, at a $350 billion valuation. Another $30 billion contingent on milestones. Up to $40 billion total.

Google Cloud is also providing five gigawatts of compute over five years through its new eighth-generation TPUs. That's a serious infrastructure commitment.

This follows Anthropic's February round and comes alongside Amazon's recent separate $5 billion commitment. Alphabet stock moved up on the news.

Here's the part worth paying attention to beyond the number.

Anthropic's uptime has been rough. Really rough. Not in a "minor inconvenience" way — in a "this falls below enterprise software minimums" way. This compute influx is the thing that could actually fix it as new Google capacity comes online.

What it means: The money matters, but five gigawatts of compute is the real headline. That's what Anthropic needs to become a reliable enterprise option — because right now, the uptime record is genuinely a problem for teams building on Claude. If reliability improves meaningfully over the next 90 days, this deal is already working.

According to Axios, Meta is laying off about 8,000 employees — roughly 10% of its entire workforce — to offset soaring AI infrastructure costs.

CapEx is forecast to rise at least 60% this year versus 2025 to fund Meta Superintelligence Labs. Free cash flow projected to fall about 83% year over year.

Welp.

This mirrors cuts happening across big tech. Amazon is planning around 16,000. Microsoft is offering buyouts to a significant chunk of staff. Block, Salesforce, and Snap have all trimmed recently.

There's another detail buried in the story. Reports surfaced that Meta may be collecting employee keystrokes to train its models. People still working at Meta might be generating training data for whatever eventually replaces parts of their job.

What it means: AI infrastructure is now expensive enough to justify cutting 10% of a company the size of Meta. That's not a one-time thing — it's a pattern across Amazon, Microsoft, Salesforce, and others all at once. The keystroke collection report is the quieter story that matters more long-term. If companies start mining employee behavior data to train models, the conversation around AI and employment gets much more complicated fast.

Also out of Google Cloud Next: Vertex AI is gone as a standalone product.

Google rebranded and expanded it into the Gemini Enterprise Agent Platform — a unified environment for building, scaling, governing, and monitoring agentic AI across enterprises.

Agent Studio brings a low-code visual interface. The Agent Development Kit handles developers going code-first. Long-running agents now have persistent memory across sessions.

Model Garden exposes over 200 models — Gemini 3.1 Pro, Lyria 3, Gemma 4, and more. Multi-agent coordination lets tasks split across specialized agents. Agents get cryptographic IDs, a centralized registry, and anomaly detection built in.

The signal is obvious. Vertex was for ML engineers. This is being positioned for the whole organization.

What it means: Google is making a direct play to pull non-technical business users into what was always a specialist platform. Agent Studio is the wedge. The governance layer — cryptographic IDs, anomaly detection, centralized registry — is what makes legal and compliance teams willing to say yes. Whether the actual user experience gets there is the open question, but the architecture to support it is now in place.

Nah. You genuinely cannot make this up.

Bloomberg reports a private Discord group accessed Anthropic's Claude Mythos Preview for more than two weeks. Not through a sophisticated attack. By guessing the URL.

One member had credentials as a third-party Anthropic contractor. That was the real entry point.

Anthropic had described Mythos as capable of autonomously finding and exploiting zero-day vulnerabilities across every major OS and browser. Access was supposed to be tightly restricted through an invite-only program called Project Glasswing.

They used it to build simple websites. And they claim access to other unreleased Anthropic models.

Anthropic says it's investigating and found no evidence of compromise beyond the third-party vendor environment. Sam Altman called the whole rollout fear-based marketing. After this, that's hard to argue with. Several GPT-5.5 benchmarks match or beat Mythos.

What it means: The perimeter failed on day one via a contractor credential and a URL guess — which raises real questions about whether the "too dangerous to release" framing was about security or about positioning ahead of an IPO. Third-party vendor environments are the weakest link in frontier AI access control, and that's an industry-wide problem, not just Anthropic's. The capability gap between Mythos and GPT-5.5 is also a lot smaller than the marketing suggested.