Feature Story
Claude's April Outage Streak Is Now a Pattern
After incidents on April 1, April 6, April 7, and April 13, Claude reliability stopped looking like isolated turbulence and started looking like an operating condition.
Coverage Index
Anthropic Coverage Desks
Desk-first discovery across governance, reliability, dependence, Claude strategy, books provenance, and leadership narrative.
68 entries
Glasswing Wasn't Just Safety. It Was Narrative Strategy.
Ben Thompson asks whether Mythos and Glasswing represent a genuine governance problem or a carefully staged danger narrative designed to justify Anthropic’s position at the frontier.
AMD Says Claude Code Regressed After the February Changes
AMD’s Stella Laurenzo publicly documented degraded reading behavior, more full-file rewrites, stop-hook violations, and weakened trust in Claude Code for serious engineering work.
Anthropic's New TPU Deal Exposed the Compute Crunch
Anthropic’s new Google and Broadcom TPU deal showed the company remains structurally dependent on whoever can supply enough compute to keep Claude running.
Recent Stress
Operational incidents, product limits, and visible pressure on daily Claude usage.
8 entries
AMD Says Claude Code Regressed After the February Changes
AMD’s Stella Laurenzo publicly documented degraded reading behavior, more full-file rewrites, stop-hook violations, and weakened trust in Claude Code for serious engineering work.
Claude Went Down Again Less Than 24 Hours Later
A second visible outage in less than a day confirmed the operational problems were not isolated incidents but a recurring service pattern.
The AI Agent Buffet Is Closed
Anthropic blocked Claude subscriptions from powering third-party agent tools like OpenClaw, converting what users thought was flat-rate access into metered usage.
Claude Code Leaks and the Security Blowback
Claude Code leaked source code and triggered security concerns during the same period of reliability complaints, compounding visible stress on Anthropic’s coding products.
Peak-Hour Limits Tightened as Demand Spiked
Anthropic reduced effective usage during peak hours even for paid plans, quietly redefining what a premium subscription actually buys.
Claude Code's Auto Mode Expanded Permission Risk
Anthropic pushed Claude Code further toward self-managed permissions just as users were already questioning its reliability and behavior on real engineering work.
Anthropic Doubled Off-Peak Limits Before Tightening the Rest
Anthropic doubled off-peak limits in a temporary March 2026 promotion, then tightened peak-hour usage and cut off third-party agent access in the weeks that followed.
Anthropic Had to Publish a Postmortem for Quality Degradation
Anthropic published a formal engineering postmortem acknowledging recurrent quality problems serious enough to require public explanation.
Governance
Safety policy, constitutional framing, and state-facing power questions.
14 entries
Does A.I. Need a Constitution?
The New Yorker asks whether Anthropic is using constitutional language to legitimize private power over systems that increasingly affect public life.
Anthropic and Alignment Put Private Sovereignty in the Open
Thompson asks whether Anthropic can simultaneously claim frontier importance and reserve final say over sovereign military use of its models.
Anthropic Dropped Its Flagship Safety Pledge
Anthropic removed its flagship safety pledge from the RSP after years of citing it as proof of superior restraint over competitors.
Palantir Sat at the Center of the Pentagon Rift
Palantir’s role in brokering Anthropic’s Pentagon access made the national-security posture operational long before the public fight over defense contracts.
Anthropic Warned Claude Could Enable Heinous Crimes
Anthropic’s own risk messaging was becoming more apocalyptic at the same time the company was deepening enterprise and government commercialization.
Anthropic’s Defense Posture Was Operational by Late 2024
The Palantir-AWS announcement shows Anthropic was already willing to place Claude inside U.S. intelligence and defense workflows, making the later Pentagon dispute a fight over limits, not participation.
Mandatory Safety Testing Was Also a Bid for Rule-Setting Power
Dario called for mandatory safety testing of frontier models, pushing regulation that would formalize the kind of evaluations Anthropic already performs.
Anthropic's 2024 RSP Update Made the Safety Framework More Flexible
The October 2024 RSP revision added implementation flexibility and broader discretion to the safety framework while preserving the formal commitment language.
U.S. Election Readiness Made Safety a Live Product Surface
Anthropic deployed election-specific controls, content policies, and public accountability measures as AI integrity became politically salient ahead of the 2024 U.S. election.
Anthropic Designed Itself to Look Unlike OpenAI
Anthropic designed its corporate structure, the LTBT, and its public governance language to differentiate itself from OpenAI’s for-profit pivot and investor conflicts.
The LTBT Was the Formal Answer to the OpenAI Split
Anthropic created the Long-Term Benefit Trust to give outside trustees veto power over safety-critical decisions, positioning the company as structurally more governable than OpenAI.
The First RSP Turned Safety Into a Formal Corporate Artifact
Anthropic’s first Responsible Scaling Policy formalized safety commitments into concrete capability thresholds and evaluation triggers, creating a measurable governance framework.
Dario’s Senate Testimony Bound Frontier AI to National-Security Logic
Dario testified before the Senate that frontier AI requires regulation, while arguing that slowing development would cede ground to geopolitical competitors and increase misuse risk.
Red Teaming Was Part of the Brand Before Claude Became Big Business
Anthropic published red-teaming research on language model harms before Claude became a commercial product, establishing the safety practice early in the company’s history.
Dependence
Cloud, capital, and partner concentration behind the independence narrative.
12 entries
Anthropic's New TPU Deal Exposed the Compute Crunch
Anthropic’s new Google and Broadcom TPU deal showed the company remains structurally dependent on whoever can supply enough compute to keep Claude running.
Anthropic's Runaway Success Is Exposing AI's Core Problem: Compute Costs
Anthropic’s rapid growth and its customer-facing usage limits are converging on the same constraint: the economics of scarce frontier compute.
Google TPU Scale Made Anthropic's Multicloud Story Less Independent
AWS remains Anthropic’s primary cloud provider, but Google’s TPU capacity is becoming too large to describe as incidental to the company’s infrastructure.
Anthropic's AWS Dependence Looked Stable Only Because Consumer Scale Wasn't the Point
Anthropic’s enterprise-first positioning changed the economics of its cloud dependence but did not remove the dependence itself.
Anthropic Told Regulators Google's Money Matters
Anthropic told regulators that Google’s ability to invest in the company was competitively important, making the dependence relationship unusually explicit in a public filing.
Anthropic Turned Claude Into a Platform With MCP
Anthropic launched the Model Context Protocol to position Claude as a platform rather than just a model, expanding its strategic footprint ahead of the later agent boom.
Amazon’s Anthropic Relationship Drew Formal Competition Scrutiny
The UK’s competition authority formally reviewed Amazon’s stake and governance rights in Anthropic, finding the relationship material enough to warrant regulatory scrutiny.
Big Tech's Anthropic Stakes Were Strategic, Not Passive
Anthropic’s major investors were also platform competitors with their own AI products and distribution agendas, making every funding round a strategic alignment.
Amazon Backed Anthropic While Building Its Own Rival Model
Amazon backed Anthropic while simultaneously building its own rival frontier model, making the partnership a strategic hedge rather than a straightforward investment.
Amazon's Anthropic Deal Was Really an AWS Strategy
Thompson identified the Amazon deal as primarily an AWS distribution play from day one, with Anthropic gaining compute and Amazon gaining a frontier model for its cloud platform.
Amazon’s Partnership With Anthropic Showed That Size Matters
Compute scarcity, not just cash, was the center of gravity in the Amazon deal, which is why the “independent alternative” story always needed qualification.
Anthropic’s Own Amazon Deal Framed Dependence as Safer Access
Anthropic framed the Amazon partnership as expanding safe AI access through cloud infrastructure, bundling its safety brand with hyperscaler compute and distribution.
Claude
Model launches, benchmark positioning, and product strategy as market narrative.
12 entries
Glasswing Wasn't Just Safety. It Was Narrative Strategy.
Ben Thompson asks whether Mythos and Glasswing represent a genuine governance problem or a carefully staged danger narrative designed to justify Anthropic’s position at the frontier.
Skyrocketing Revenue Raised the Cost of Every Principle
Anthropic’s enterprise revenue was accelerating rapidly, raising the economic stakes behind every subsequent policy softening and government compromise.
Claude's Real Business Model: Enterprise Coding, Not Consumer Love
Anthropic’s revenue increasingly depends on coding tools and enterprise contracts, making Claude’s consumer identity secondary to its business positioning.
Anthropic Played the Benchmark Game While Critiquing It
Anthropic revoked OpenAI’s access to Claude over benchmarking practices while the same kind of rival testing remains standard across the frontier-model market.
Claude 4 Made the Agents Strategy Explicit
Claude 4 made agents and coding workflows the center of Anthropic’s product strategy, tying revenue growth directly to the reliability of those tools.
Computer Use Turned Claude From Assistant Into Agent
Anthropic’s own launch language made the ambition explicit: Claude was no longer just a chatbot but a system meant to operate software, take actions, and justify a more agentic premium.
Anthropic's Computer Agent Claims Arrived With Verification Gaps
Anthropic’s computer-use benchmarks were promising but not independently verified, and Claude’s actual performance remained far from human reliability on real tasks.
Dario Framed Anthropic as an Underdog While Defending Frontier Scale
Dario openly tied concentration, state capacity, inequality, and scaling pressure to Anthropic’s competitive position while framing the company as an underdog against larger rivals.
Claude 3.5 Sonnet Deepened the Coding-and-Enterprise Story
Anthropic’s 3.5 Sonnet launch sharpened the product identity that later drove revenue: faster iteration, stronger coding, and benchmark-led positioning aimed squarely at daily work.
Claude 3 Turned Benchmark Leadership Into the Product Pitch
Claude 3 launched with explicit claims of benchmark supremacy across reasoning, math, and coding, pairing safety branding with direct performance competition against GPT-4.
Claude 2.1 Tied Bigger Context Windows to Enterprise Trust
Anthropic sold Claude 2.1 on a now-familiar bundle: larger context, lower hallucination rates, tool use, and pricing changes framed as reliability for enterprise buyers.
Claude 2 Paired Public Access With Exam-Score Marketing
Claude 2 launched with public access, frontier benchmark scores, longer context, improved coding, and safety claims bundled together for the first time.
Books and Provenance
Copyright disputes, training-data provenance, and litigation milestones.
9 entries
The Books Record Spilled Into a New Music Piracy Front
The books case did not end the legal overhang; it created a template for a broader music-side piracy story built on the same provenance concerns.
The Fair-Use Win Still Ended in Settlement
Anthropic’s fair-use position survived judicial review, but the piracy-based claims were serious enough to resolve with a landmark settlement.
Napster-Style Downloads Entered the Judicial Record
The class-certification order describes Anthropic’s book-acquisition behavior as deliberate and scalable, with downloads resembling Napster-era mass piracy.
Authors Framed the Ruling as Fair Use Plus Mass Piracy
The Authors Guild framed the split ruling as a partial win on fair use but a confirmation that mass piracy of copyrighted works remained a live legal problem.
Judge Alsup Split the Case in Two: Training OK, Theft Not OK
Judge Alsup ruled that training on copyrighted works could qualify as fair use, but the separate claims about acquiring books from pirate libraries survived.
The Bartz Complaint Put the Books Record Into Public Litigation
The authors’ complaint directly alleged large-scale use of copyrighted and pirated books to train Claude, putting the provenance question into public litigation.
Anthropic Told the Copyright Office Why AI Training Should Be Fair Use
Anthropic argued to the U.S. Copyright Office that AI training on copyrighted material should qualify as fair use, establishing its legal position before the books lawsuits began.
Books3 Made the Training-Data Fight Concrete
Once Books3 became searchable, the training-data dispute stopped being abstract. Authors could identify titles, and the provenance problem became publicly legible.
Pirated Books Were Identifiable Long Before the Lawsuits Matured
The Atlantic traced pirated books in the Books3 dataset to specific titles and authors months before lawsuits forced AI companies to answer for their training data sources.
Dario
Founding claims, leadership rhetoric, and scaling doctrine around Amodei.
12 entries
What Is Claude? Anthropic Doesn't Know, Either.
The New Yorker profiles Anthropic as an interpretability lab forced into commercialization, struggling to reconcile the two identities.
Machines of Loving Grace Shifted the Tone From Fire Alarm to Salvation
Dario reframed AI not only as dangerous and urgent but as the engine of a compressed century of human progress, marking a public pivot from alarm to optimism.
Among the A.I. Doomsayers
The New Yorker profiled the AI safety community’s apocalyptic culture and social dynamics, providing the ideological backdrop for Anthropic’s moral positioning and fundraising success.
Why Dario Quit OpenAI Became the Anthropic Origin Story
Dario told Fortune he left OpenAI over safety disagreements and misaligned incentives, establishing the founding narrative that Anthropic needed to exist as a more responsible alternative.
Dario's Scaling Maximalism Was Public by 2023
When Dario said he was not sure there were limits to AI, he made explicit the scaling worldview that sits underneath both Anthropic’s safety warnings and its drive toward larger systems.
Claude's Constitution Made Anthropic's Value Claims Explicit
Anthropic published Claude’s constitution as an explicit set of values and principles governing its behavior, making the company’s alignment approach publicly legible.
Introducing Claude Turned Helpful-Honest-Harmless Into a Product Story
Claude’s first broad release is where Anthropic’s research identity became a commercial interface, packaging steerability and harmlessness as reasons to choose its assistant over competitors.
Core Views on AI Safety Became the Company Doctrine
Anthropic laid out its core position: AI scale is inevitable, catastrophic risk is real, and the responsible path is to build frontier systems while actively managing their dangers.
Constitutional AI Gave Claude a Distinct Moral Operating System
The Constitutional AI paper described a method for training Claude using written principles rather than human feedback alone, giving Anthropic a technical basis for its alignment claims.
Series B Joined Safety Language to Compute Scale
The Series B announcement already contains the full Anthropic formula: more capital, more infrastructure, more scale, and a promise that safety research will justify all three.
Anthropic Started as a Safety Story. The Question Is What Survived Contact With Scale.
Anthropic raised $124 million to build reliable, steerable AI systems, grounding the company’s founding promise in safety research and general-purpose capability.
Scaling Laws Supplied the Technical Worldview Behind Anthropic
The scaling-laws paper established that model performance improves predictably with compute, data, and parameters, supplying the technical premise that larger models would be both more capable and more dangerous.