This Week in AI: The Mythos Effect, $200B for Google, and an FDA for Models

Five business days. That is how long it took for one Anthropic model release to drag the U.S. government, the three biggest cloud providers, and the OpenAI product roadmap into a new shape. If you have been wondering when AI policy was going to stop being a slide deck and start being a contract, this was the week.

The unifying thread is Mythos. Anthropic shipped a model that finds and exploits network vulnerabilities; Washington noticed; everything else this week is a reaction to that one fact. Here is the full roundup.

1. CAISI Got the Keys to Google, Microsoft, and xAI

On Tuesday May 5, the Center for AI Standards and Innovation (CAISI), housed inside the Department of Commerce, announced pre-deployment evaluation agreements with Google DeepMind, Microsoft, and Elon Musk's xAI. The agreements give CAISI access to test models in classified environments before they ship publicly, covering cybersecurity, biosecurity, and chemical-weapons risk surfaces.

OpenAI and Anthropic signed equivalent agreements back in 2024 under the Biden administration. Those got renegotiated this week to align with Commerce Secretary Howard Lutnick's directives and the Trump White House's AI Action Plan. Net effect: every U.S. frontier lab is now subject to the same pre-launch evaluation regime.

CAISI says it has already completed more than 40 model evaluations, some of them on models that have not been released. The America First Policy Institute, despite supporting the mission, called CAISI "chronically underfunded" with roughly 30 staff and $30M since 2024. The capability is in place; the headcount is not. Expect that gap to be the fight inside the executive order that's allegedly being drafted next.

2. The "FDA for AI" Executive Order

National Economic Council Director Kevin Hassett confirmed on Tuesday that the White House is studying an executive order that would create "a clear road map" for how advanced AI systems get evaluated before release. The framing he reached for: FDA drug approval. New model goes through pre-market review, gets a verdict, ships or does not.

Two things are notable about this. First, it is a real reversal: the same administration that spent the first half of 2025 dismantling Biden-era AI regulations is now drafting a mandatory pre-launch vetting regime. Second, the trigger was named in the reporting: Anthropic's Mythos. A model that can find network vulnerabilities at scale convinced the people in the room that voluntary commitments were not going to cut it.

We covered Mythos when it dropped in our analysis of what Mythos actually does and Anthropic's decision to ship it anyway in Claude Mythos, Not Afraid. The product policy and the federal policy are now linked.

3. Anthropic Promised Google $200 Billion

The Information reported Tuesday that Anthropic has committed to spending $200 billion with Google Cloud over five years for compute and Broadcom-built TPU capacity, with the new gigawatt-scale capacity expected to come online starting in 2027. Both companies declined to confirm; nobody is denying the number.

Some scale: Alphabet disclosed roughly $400B of revenue backlog last week. If The Information's read holds, Anthropic is more than 40% of it. Add OpenAI's contracts at AWS, Azure, and (now) Google, and two private companies account for over half of the $2 trillion in backlog the three big U.S. cloud providers are sitting on.

Read this alongside Alphabet's earlier $40B Anthropic equity stake and you see what is actually happening: Google is bankrolling its own most credible model competitor, then re-collecting most of the money on the compute side. The hyperscaler-frontier-lab relationship has stopped looking like a customer relationship and started looking like vertical integration with extra steps.

4. OpenAI Shipped GPT-5.5-Cyber

On Wednesday May 7, OpenAI rolled out GPT-5.5-Cyber, a variant of its flagship tuned for vulnerability identification, triage, patch validation, and malware analysis. Limited preview, vetted teams only. Roughly one month after Anthropic's Mythos.

The product framing is defensive: helping security teams move faster on the same surface Mythos can attack. The market framing is the one that matters. OpenAI just confirmed, with a SKU, that cyber capability is now a category every frontier lab needs a credible answer in. Expect a Google version inside a month.

For context on why this is now a model class instead of a feature, our base GPT-5.5 launch breakdown covers the underlying capability gain that makes a Cyber tier viable in the first place.

5. Cohere Bought Aleph Alpha. Sovereign AI Got a Vendor.

Last week Cohere closed its $20B-valuation merger with Germany's Aleph Alpha, with €500M in structured financing led by Schwarz Group. The Canadian and German digital ministers showed up in Berlin to bless it personally, citing the Canada-Germany Sovereign Technology Alliance signed earlier in 2026.

The thesis is simple. Regulated sectors (defense, finance, healthcare) and European public sector buyers want a frontier-tier model they can run inside their own jurisdiction with no Microsoft, Google, or AWS in the data path. Cohere brings the LLM stack and enterprise sales motion; Aleph Alpha brings the European-language coverage and the regulatory relationships. The combined entity is the first credible North-Atlantic answer to the U.S.-China duopoly on frontier compute.

Pair this with SAP's Prior Labs acquisition and Europe now has two distinct sovereign-AI plays inside two weeks: tabular foundation models from SAP, sovereign LLMs from Cohere/Aleph Alpha. Different product surfaces, same political tailwind.

6. China Blocked Meta's Manus Acquisition

Beijing's National Development and Reform Commission formally prohibited Meta's $2B acquisition of Chinese agent startup Manus this week. First state-level block of an inbound AI acquisition by China. The signal: Chinese AI companies are now classified as strategic enough that a U.S. tech giant cannot buy one, even at a substantial premium.

Functionally this is the mirror image of the U.S. CFIUS regime that has blocked Chinese buyers for years. AI is now a sovereign asset class on both sides of the Pacific. Expect more of these.

7. Google Quietly Shipped Gemini 3.1 Flash-Lite

Easy to miss inside a policy week, but worth the mention. Google introduced Gemini 3.1 Flash-Lite this week at $0.25 per million input tokens, with claimed 2.5x faster response times and 45% faster output generation than the previous Flash tier. The cheap end of the market keeps falling, exactly as our pricing floor analysis said it would.

The two halves of the market are now visibly diverging: cyber, agentic, and frontier work on $5+ input tiers; commodity inference racing toward zero on Flash, Mini, and Nano SKUs. Same week, both directions, neither contradicting the other.

Our Take

The story of the week is not any single drop. It is that the frontier finally accepted that capability is going to come with a regulator. The CAISI agreements, the executive order in draft, GPT-5.5-Cyber, the $200B Google contract: these are all the same trade, priced into different documents. If you can find network vulnerabilities at scale, you are a defense system; defense systems get vetted; vetting needs compute; compute needs contracts.

Practical implications for builders. One, assume your frontier model provider will be subject to a pre-launch evaluation regime by the end of 2026. Build for that, not against it. Two, the cyber capability tier is now a real product category. If you ship security software, you have an OpenAI and an Anthropic SKU coming for your seat budget. Three, the policy moat just became a real moat. U.S.-aligned labs now have something Chinese labs structurally cannot get: classified-environment evaluation that doubles as a federal procurement signal.

We will be tracking the executive order, the CAISI funding fight, and the next cyber-tier model release. The chart that matters next week is whether Google ships its own response to GPT-5.5-Cyber. If they do, the cyber category is real. If they sit it out, the capability surface is narrower than the policy reaction suggests. Either way, we will know shortly.

See you Monday.