AI Briefing — Week of May 18, 2026

This week's signals coalesce around one company: Anthropic simultaneously filed for a near-trillion-dollar valuation, shipped products targeting small businesses and Big Law, released a cybersecurity model it deemed too dangerous to publish widely — and that model's existence nudged the Trump administration toward its first real AI oversight posture. Meanwhile, OpenAI moved Codex to mobile, Google pre-staged Gemini for Android ahead of I/O, a Fields medalist publicly credited GPT-5.5 Pro with PhD-level mathematics, and Isomorphic Labs closed the largest private AI drug discovery round in history. Here is what you need to know before markets open.

Products

OpenAI brings Codex to ChatGPT mobile

On May 14, OpenAI released a ChatGPT mobile update that lets users monitor, steer, and approve Codex coding tasks from an iPhone, iPad, or Android device 1. Codex now has 4 million weekly users 2. The update also ships Remote SSH for enterprise environments, programmatic access tokens (Enterprise/Business tier only), Hooks for custom repo-level behavior, and HIPAA compliance for qualifying enterprise workspaces using Codex locally. The practical unlock for founders: a developer can kick off a multi-hour task from a laptop, then approve sub-tasks or redirect the agent from a phone during transit.

Google repositions Gemini as the Android OS layer

Ahead of Google I/O (May 19), Google announced Gemini Intelligence — a redesign of Android that lets Gemini understand screen context and run multi-step tasks across apps 3. Demonstrated capabilities include extracting a guest list from Gmail, building a grocery cart in Instacart, and handing back control to the user before checkout. The feature rolls out this summer to Samsung Galaxy and Google Pixel first, then extends to Android watches, cars, glasses, and laptops. Google is also shipping a redesigned Android Auto and an "AI-native" Gemini-first laptop line called Googlebook 4.

Anthropic launches Claude for Small Business and Claude for Legal

On May 13, Anthropic released Claude for Small Business, built on top of its Claude Cowork agent framework 5. The package integrates with QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365, and ships with 15 ready-to-run workflows covering finance, HR, sales, and customer service. On May 12 — one day earlier — Anthropic launched Claude for Legal, a suite of 12 new tools designed to place Claude directly in law firm workflows rather than operating only as a layer beneath specialist vendors like Harvey or Legora 6. The vertical push signals Anthropic shifting from infrastructure provider to direct enterprise software competitor.

Models

Claude Opus 4.7 gets a 2.5× fast mode — at 6× the price

Anthropic released Fast Mode for Claude Opus 4.7 in research preview on May 18, delivering 2.5× faster output token generation with the same model weights 7. Pricing jumps to $30/$150 per million input/output tokens — roughly 6× the standard Opus 4.7 rate. Available in the API, Claude Code, and Cursor. The tradeoff worth sizing up: at that premium, Fast Mode only makes economic sense for latency-critical agentic pipelines where wall-clock time directly converts to revenue (e.g., time-sensitive trading workflows or real-time customer-facing agents).

Fields medalist says GPT-5.5 Pro produced PhD-level number theory in an hour

Cambridge mathematician and Fields medalist Timothy Gowers posted that GPT-5.5 Pro solved a set of open problems from a recent additive number theory paper within roughly one hour, turning an exponential bound into a polynomial result 4. A customized GPT-5.5 variant also produced a Lean-verified new proof related to off-diagonal Ramsey number asymptotics. Gowers wrote that "we will face a crisis very soon" in PhD student recruitment as a result. OpenAI simultaneously split its model access tiers, creating a separate entry point for AI-assisted cyber defense called Daybreak — a bundle of GPT-5.5 and Codex Security oriented toward vulnerability discovery and patch generation.

Google's Gemma 4 adds speculative decoding — up to 3× inference speedup

Google released a multi-token prediction draft module for Gemma 4 that uses speculative decoding to speed inference up to 3× with no output quality degradation 4. Particularly relevant for local and edge deployments — the module is already available to community builders running Gemma 4 on-device.

Anthropic's Mythos: the model they won't release

Mythos, a cybersecurity-focused model announced last month, continued to generate news without shipping publicly. Anthropic shared access with approximately 40 organizations managing critical infrastructure, including Mozilla, which used it to identify and patch 271 vulnerabilities in Firefox 150 8. An open-source review of Mythos against the curl codebase was less convincing — only 1 of 5 claimed vulnerabilities was confirmed real and low-severity. Anthropic also agreed to brief the Financial Stability Board on systemic cyber weaknesses in global financial infrastructure that Mythos identified 9. U.S. banks were separately reported to be rushing to patch IT weaknesses flagged by Mythos assessments.

Funding

Anthropic in talks to raise at a $950B valuation

Anthropic is in discussions to raise $30–$50 billion in new funding at a valuation of up to $950 billion — 2.5× its $380 billion valuation from just three months ago and above OpenAI's March valuation of $852 billion 10. CEO Dario Amodei recently said the company had reached a $30 billion revenue run rate with 80× growth expected this year. The talks are ongoing and could still fall apart, per the NYT's sources.

Isomorphic Labs closes $2.1B Series B — the largest private AI drug discovery deal on record

Alphabet-backed AI drug design company Isomorphic Labs raised $2.1 billion in a Series B led by Thrive Capital, with participation from Alphabet/GV, MGX, Temasek, CapitalG, and the U.K. government's AI fund 11. The round follows a $600M Series A a year ago, bringing total funding to $2.7 billion. Isomorphic uses models including AlphaFold 3 to design potential drug molecules; its IsoDDE engine already has commercial deals with Eli Lilly and Novartis. Founder Demis Hassabis described the goal as building a drug design engine "at scale, driving us forward in our mission to solve all disease."

Exaforce raises $125M Series B for AI-native cybersecurity SOC

Exaforce, a three-year-old startup that uses AI agents ("Exabots") to automate security operations, raised $125 million at a $725 million valuation in a round backed by HarbourVest, Peak XV, Mayfield, Khosla Ventures, and Seligman Ventures 12. The raise follows a $75M Series A from last year, bringing total funding to $200M. Exaforce claims its platform reduces manual security analyst work by up to 90% and has added 20 customers including Replit and Guardant Health since going to market in Q4 2025. The company recently launched "vibe hunting" — natural language queries for threat investigation.

Monaco secures $50M Series B for AI-native sales

Monaco, an AI-native sales platform, raised a $50 million Series B led by Benchmark 13. The San Francisco company targets revenue teams with AI-driven workflows and cited strong customer demand as the driver.

Regulation & compliance

Trump administration signals a shift on AI safety

In what appears to be a meaningful departure from 18 months of deregulatory rhetoric, the White House is now considering an executive order that would create an AI working group of tech executives and government officials to examine oversight procedures — including a possible formal government review process for new models 14. The proximate cause, per NPR reporting: the Anthropic Mythos release jolted officials into using the word "safety" after months of replacing it with "security." Treasury Secretary Scott Bessent said on Fox News: "There is a very important calculus here between innovation and safety." Brookings technology policy researcher Kyle Chan described the shift as "a dramatic reversal from just a year ago." White House officials have said any policy announcement would come directly from the president; Congress has shown no movement toward formal AI legislation.

An internal snag: an AI executive action stalled on May 13 due to White House infighting between officials prioritizing innovation and those pushing for guardrails 15.

Colorado's SB 189 replaces its landmark AI bias law with a lighter disclosure bill

Colorado Governor Jared Polis signed Senate Bill 189 on May 12, replacing the state's more comprehensive 2024 law that had sought to prevent algorithmic discrimination 16. The new law retains AI disclosure requirements but removes the high-risk system liability framework that had alarmed enterprise tech vendors. Colorado had been among the first U.S. states to pass substantive AI legislation; this reversal reflects broader pressure from the Trump administration, which has conditioned federal broadband funding on states not creating AI patchwork regulations 17.

Trump-Xi summit puts AI chip exports at the center of U.S.-China negotiations

President Trump's Beijing visit this week made AI chip exports and supply chain security central agenda items 18. The U.S.-China trade truce did not resolve the gallium and germanium export restrictions — materials China controls and the U.S. needs for semiconductor manufacturing — meaning the chip supply chain remains a live risk variable for any AI infrastructure play with China exposure.

The week in one line: Anthropic is now valued closer to $1 trillion than to $100 billion — and the model it decided not to release may be doing more to reshape U.S. AI policy than any of the models it did.