← All issuesSaturday, May 16, 2026Vol. I · No. 01
The
AI Gazette
Weekly Digest"All the AI that's fit to print."Founded in Switzerland
Lead · A Week in Artificial Intelligence

The labs turned on each other.

Anthropic leased the cluster that trained Grok. OpenAI paid Anthropic's customers to defect. Salesforce opened the doors to the machines. One unusually loud week.

The polite phase is over. Anthropic raised prices on its heaviest users — and within hours, OpenAI offered them two free months to leave. Anthropic also brushed a $900 billion valuation. The boldest move of the week, though, may have come from Salesforce, which quietly handed its entire platform to AI agents. Twelve stories, inside.

Inside This Issue

  1. Anthropic is leasing Grok's data center to double users' limits.p. 2
  2. Heavy Claude users just got a quiet price hike.p. 3
  3. Anthropic is now worth $900 billion. IPO eyed for October.p. 4
  4. Claude now plugs into the tools small businesses already use.p. 5
  5. OpenAI is paying companies to leave Claude.p. 6
  6. On the main coding test, OpenAI and Claude are basically tied.p. 7
  7. A free AI that rivals ChatGPT — with no content filters.p. 8
  8. Hermes can now spin up a full coding sandbox on demand.p. 9
  9. Salesforce just made itself usable by AI agents — not just humans.p. 10
  10. The head of Y Combinator gave away his personal "second brain."p. 11
  11. The smartest AI bet isn't always the AI itself.p. 12
  12. Only 9% of companies have actually replaced a job with AI.p. 13
Page 2The Compute Race · Anthropic2026-05-16
From Anthropic.com · 6 May 2026

Anthropic is leasing Grok's data center to double users' limits.

In a stark sign that computing power, not model design, is the binding constraint, the firm took over a rival's cluster to push limits higher.

— Special to The AI Gazette —
Anthropic's announcement, "Higher usage limits for Claude and a compute deal with SpaceX."
Anthropic's announcement, "Higher usage limits for Claude and a compute deal with SpaceX."anthropic.com/news/higher-limits-spacex

Anthropic took over the data center xAI used to train Grok — its own competitor — and is using that firepower to permanently double how much Claude users can do each day. Weekly caps also rise by half through mid-July.

Why It
Matters

Compute, not the model itself, is the real bottleneck in AI right now. Anthropic just locked in the capacity to keep raising limits.

Page 2The AI Gazette · 2026-05-16Read the source →
Page 3Markets · Anthropic2026-05-16
From The New Stack · 14 May 2026

Heavy Claude users just got a quiet price hike.

By splitting automated workloads out of its $100–$200 subscriptions, Anthropic raises bills on its most demanding customers without changing the sticker.

— Markets Desk —
The New Stack: "Anthropic splits billing again: Agent SDK gets separate credit pools."
The New Stack: "Anthropic splits billing again: Agent SDK gets separate credit pools."thenewstack.io/anthropic-agent-sdk-credits

Power users had been running scripts and automations through their $100–$200/month Claude subscription — basically treating it like unlimited access. From 15 June, that usage gets carved off and billed at API rates instead. Much, much more expensive.

Why It
Matters

Anthropic raised prices on its heaviest users without touching the sticker. If anyone on your team automates anything through Claude, the bill is about to jump.

Page 3The AI Gazette · 2026-05-16Read the source →
Page 4Capital Markets · Anthropic2026-05-16
From Bloomberg · 12 May 2026

Anthropic is now worth $900 billion. IPO eyed for October.

A fresh $30 billion round more than doubles the firm's valuation in three months and sets up what may be the largest technology listing of the decade.

— Capital Markets Desk —
Anthropic's earlier Series G announcement — already eclipsed by the new round.
Anthropic's earlier Series G announcement — already eclipsed by the new round.bloomberg.com · Anthropic in talks to raise $30B at $900B valuation

Anthropic is raising another $30 billion at a $900 billion valuation — more than doubled in three months. An October IPO is reportedly on the table.

Why It
Matters

Anthropic now ranks as the second-biggest AI company in the world. The recent pricing changes suddenly make sense — they're tuning the business for public markets.

Page 4The AI Gazette · 2026-05-16Read the source →
Page 5Small Business · Anthropic2026-05-16
From Anthropic.com · 13 May 2026

Claude now plugs into the tools small businesses already use.

A free package of prebuilt integrations wires Anthropic's assistant into QuickBooks, PayPal, HubSpot and Google Workspace — automating tasks and orchestrating workflows across the apps SMBs already run.

— Small Business Desk —
"Introducing Claude for Small Business" — Anthropic's new SMB push.
"Introducing Claude for Small Business" — Anthropic's new SMB push.anthropic.com/news/claude-for-small-business

Anthropic shipped a free package of prebuilt integrations: Claude now connects directly into QuickBooks, PayPal, HubSpot and Google Workspace, automating routine tasks and managing workflows across the tools small businesses already run.

Why It
Matters

Anthropic isn't replacing small-business software — it's becoming the workflow layer on top of it. The value left for advisors and boutiques is industry-specific orchestration, not generic plumbing.

Page 5The AI Gazette · 2026-05-16Read the source →
Page 6Markets · OpenAI2026-05-16
Via @sama on X · 13 May 2026

OpenAI is paying companies to leave Claude.

Within hours of Anthropic's pricing shake-up, Sam Altman offered two free months of Codex to any company willing to switch sides.

— Telegraphed from San Francisco —
The chief executive's announcement on X, drawing more than 21,000 endorsements.
The chief executive's announcement on X, drawing more than 21,000 endorsements.x.com/sama/status/2054626219858293128

For the next 30 days, OpenAI is giving away two months free of its coding tool — but only to companies switching over from a rival (read: Anthropic).

Why It
Matters

A direct shot at Anthropic, hours after their pricing change. Expect every engineering team you know to be running side-by-side trials this month.

Page 6The AI Gazette · 2026-05-16Read the source →
Page 7Research · Benchmarks2026-05-16
From llm-stats.com · May 2026

On the main coding test, OpenAI and Claude are basically tied.

New scores put the two leading models within a single point of each other. Even OpenAI now admits the headline number is unreliable.

— Research Desk —
The SWE-bench Verified leaderboard, dominated by the two largest American AI laboratories.
The SWE-bench Verified leaderboard, dominated by the two largest American AI laboratories.llm-stats.com/benchmarks/swe-bench-verified

New scores put OpenAI at 88.7% and Anthropic at 87.6% on the most-cited coding test. Each wins on different sub-tests. Even OpenAI admits the headline number is no longer reliable.

Why It
Matters

The "which AI is smartest" question is over. The real question is which one fits how your team actually works.

Page 7The AI Gazette · 2026-05-16Read the source →
Page 8Open Source · Nous Research2026-05-16
From Nous Research

A free AI that rivals ChatGPT — with no content filters.

Nous Research released a downloadable model that matches the frontier labs on reasoning, and will answer questions the major American firms refuse to touch.

— Open Source Desk —
"Hermes 4" — Nous Research's open-weights model, in three sizes up to 405 billion parameters.
"Hermes 4" — Nous Research's open-weights model, in three sizes up to 405 billion parameters.hermes4.nousresearch.com

Nous Research released Hermes 4: a free, downloadable AI that matches OpenAI and Anthropic on reasoning tests, and will answer questions the big labs refuse to touch.

Why It
Matters

For companies that can't send sensitive data to outside AI providers — banks, hospitals, law firms — there's now a credible model they can run on their own servers.

Page 8The AI Gazette · 2026-05-16Read the source →
Page 9Open Source · Orchestration2026-05-16
From hermes-agent.nousresearch.com

Hermes can now spin up a full coding sandbox on demand.

A new skill lets the open-source agent wrap OpenAI's coding tool inside its own sandboxed environment — a complete workstation on demand.

— Technology Desk —
The Codex skill's metadata — bundled with Hermes, distributed under the MIT licence.
The Codex skill's metadata — bundled with Hermes, distributed under the MIT licence.hermes-agent.nousresearch.com · skills/codex

Open-source agent Hermes added a skill that wraps OpenAI's coding tool, Codex, inside its own sandboxed environment — terminal, files, the lot. Hermes can now spin up a full coding workstation on demand.

Why It
Matters

Agents are starting to use each other as tools. Hermes doesn't need to be good at coding — it just needs to know when to hand the job to Codex.

Page 9The AI Gazette · 2026-05-16Read the source →
Page 10Enterprise · Salesforce2026-05-16
From VentureBeat · TDX 2026, San Francisco

Salesforce just made itself usable by AI agents — not just humans.

Every action in the world's largest customer-software platform is now callable by an AI agent — no browser, no clicks required.

— Enterprise Desk, San Francisco —
The TDX 2026 announcement: the most ambitious architectural rebuild in Salesforce's 27-year history.
The TDX 2026 announcement: the most ambitious architectural rebuild in Salesforce's 27-year history.venturebeat.com · Salesforce launches Headless 360

Every action you'd normally do inside Salesforce — create a deal, update a contact, run a report — can now be done by an AI assistant directly, without anyone opening the app.

Why It
Matters

The largest business-software company on the planet just declared itself AI-native. Expect Microsoft, SAP and Workday to follow within a year.

Page 10The AI Gazette · 2026-05-16Read the source →
Page 11Tools · Y Combinator2026-05-16
Via @garrytan on X · 10 April 2026

The head of Y Combinator gave away his personal "second brain."

Garry Tan published the setup he uses to keep an AI agent's memory durable and clean across sessions — and gave it away under a permissive licence.

— Filed from Mountain View —
Mr. Tan announcing "GBrain" on X — the system, he writes, is "exactly my OpenClaw/Hermes Agent setup."
Mr. Tan announcing "GBrain" on X — the system, he writes, is "exactly my OpenClaw/Hermes Agent setup."x.com/garrytan/status/2042497872114090069

GBrain is a three-layer memory system: Signal Detection flags what's worth remembering, Brain Ops keeps the knowledge graph clean overnight, and Search retrieves context with hybrid vector + keyword lookup. Open source, MIT-licensed.

Why It
Matters

Persistent, structured memory is becoming the real moat in AI workflows. Vendor-managed chat history forgets; an owned three-layer brain compounds. Worth borrowing the blueprint instead of designing one from scratch.

Page 11The AI Gazette · 2026-05-16Read the source →
Page 12Opinion · Theory Ventures2026-05-16
From tomtunguz.com · 15 May 2026

The smartest AI bet isn't always the AI itself.

A note from Tomasz Tunguz argues that the businesses sitting one layer beneath the great models are compounding faster than the labs themselves.

— Editorial Page —
Mr. Tunguz's essay, "The First Derivative of Inference."
Mr. Tunguz's essay, "The First Derivative of Inference."tomtunguz.com · The First Derivative of Inference

The best AI businesses either sell AI directly (Anthropic, OpenAI, ElevenLabs) or ride its coattails. Every voice agent built on ElevenLabs needs phone numbers — so Twilio sells more. Every AI call needs monitoring — so Datadog grows too.

Why It
Matters

When LLM usage goes up, a long list of adjacent businesses goes up with it. Often a quieter, safer bet than building the model itself.

Page 12The AI Gazette · 2026-05-16Read the source →
Page 13Economy · Labour Markets2026-05-16
From CNN Business · 10 May 2026

Only 9% of companies have actually replaced a job with AI.

The headlines say artificial intelligence is wiping out work. The data tell a quieter story: hiring is slowing, but replacement remains rare.

— Labour Correspondent —
The view of an empty office — but, the data suggest, not for the reasons the headlines claim.
The view of an empty office — but, the data suggest, not for the reasons the headlines claim.cnn.com · AI isn't actually 'taking' your job

The headlines say AI is wiping out jobs. The data says different: just 9% of companies have actually replaced a role. The World Economic Forum still forecasts a net gain of 78 million jobs by 2030. What's really happening is slower hiring, not layoffs.

Why It
Matters

Position AI to your team as "do the work you can't currently hire for" — not "fire people." That framing actually lands.

Page 13The AI Gazette · 2026-05-16Read the source →

Where does AI actually fit in your business?

Our AI Audit gives you the answer — and gets you started in under a week. No retainer. No slide decks. A clear plan, and the first workflow live.

Book a 30-min call →