AI News from THE DECODER

Latest coverage from THE DECODER, summarized and scored for signal.

• AI labs are failing to keep their own systems in check — No AI company fully implements basic safety controls for its own internal AI systems, according to Guidelight's first independent scorecard — Anthropic and OpenAI earn C+, Google D+, xAI D−, and Meta an outright F on six core safety practices.
• Anthropic says any lab can now let a language model agent run the whole protein design stack — Anthropic's Claude models autonomously ran a full protein design pipeline — installing and orchestrating existing open-source biology tools — achieving a 26.8% binding hit rate on novel minibinders, nearly double the industry benchmark of 10–15%, though independent replication is still pending.
• Anthropic passes OpenAI on revenue for the first time — Anthropic's quarterly revenue hit $11.6B — surpassing OpenAI's $6.7B for the first time — as Claude Code adoption drives a sevenfold year-over-year increase in Anthropic's annualized revenue rate to $65B while OpenAI's operating margin stays negative ahead of an expected IPO.
• OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous — OpenAI paused reinforcement learning on its Astra model after it crossed a cyber-critical capability threshold — the ability to enable sophisticated cyberattacks at scale. The company now devotes 20% of inference compute to behavioral monitoring and deploys AI agents to investigate other AI agents for dangerous behavior.
• New benchmark ranks search APIs for AI agents on quality, cost, and speed — Artificial Analysis released a Search Index benchmarking 7 search API providers for AI agents: Parallel scores 75, Exa 74, Firecrawl 73 on a composite of quality, cost, and speed — and better search reduces total agent cost by cutting downstream token use by 40%+.
• Anthropic CEO says AI centralizes by nature and open models just shift power to whoever owns the chips — A public X debate between Dario Amodei, Yann LeCun, and investor Gavin Baker exposes a core AI policy fault line: Amodei argues open models shift power to compute-rich players rather than end users, while LeCun compares open AI to the printing press.
• As AI beats doctors, regulators shouldn't force a human into the loop, JAMA piece says — A JAMA opinion piece co-authored by a top bioethicist and a health AI CEO argues regulators should not mandate human oversight of medical AI, citing studies showing autonomous AI now matches or outperforms doctors on five core medical reasoning tasks — though nearly all evidence comes from simulations, not actual patient care.
• DOJ probes Andreessen Horowitz over partners sitting on competing AI boards — The US Justice Department is investigating Andreessen Horowitz for a potential Clayton Act violation: cofounder Ben Horowitz sits on the Databricks board while partner Martin Casado sits on the Fivetran board — two competing AI data companies both backed by the firm.
• OpenAI launches a ChatGPT version built for teens — OpenAI is rolling out a dedicated teen mode for ChatGPT that applies stricter safeguards for users 13–17, using 2,000+ behavioral signals rather than direct age verification — a response to mounting lawsuits including Florida's state suit over harmful content exposure to minors.
• Anthropic's per-token cost runs 4.4 times the average on Vercel, and developers keep paying — Vercel's July AI Gateway data reveals Anthropic captured 65.1% of spending while handling only 30% of tokens — its per-token cost runs 4.4x the platform average. Claude Fable 5 claims 13.2% of spend with 9-in-10 teams being new customers, signaling premium models retain pricing power even as average token prices fell 13.6% industry-wide.
• Claude Code gets a /design command that lets developers create UI mockups right in the terminal — Anthropic released an early preview of the /design command in Claude Code, letting developers generate UI mockups as artboards in the terminal by typing "/design a few options for {feature}" — Claude reads the existing codebase, matches the current UI style, and produces shareable Artifacts before any code is written.
• AI systems quietly drop user instructions when they compress context — Penn State researchers found AI context compression (compaction) drops 83% of user-set session rules on average — leaving only 17% intact — creating silent safety failures; a Qwen3.5-9B add-on module preserves over 90% of these constraints as a plug-and-play fix.
• Anthropic increases revenue sevenfold, hits annualized rate above $65 billion — Anthropic's annualized revenue topped $65 billion in July 2026—a 7x increase year-over-year—with a $1 trillion IPO potentially coming as early as fall 2026, which would make it among the most valuable companies ever to go public.
• AirTag reveals how Amazon destroys rare books for AI training — Amazon's warehouse team physically destroys books after scanning their contents for Nova model training data, confirmed by 404 Media via an AirTag planted in a rare-book shipment; Anthropic ran an identical operation called 'Project Panama' which a court ruled fair use.
• OpenAI signs record Ohio data center lease with Nvidia backing up to $105 billion — OpenAI signed a 20-year lease for an 8-gigawatt Ohio data center with Nvidia guaranteeing up to $105 billion in residual value and locking in as exclusive chip supplier — the largest single data center commitment announced to date.
• AI video market has bounced back from Sora's false start — The AI video market has matured into a financed industry: Higgsfield raised $400M at a $5.4B valuation (up from $1.3B eight months ago), Promise is filming Hollywood horror films using real-time AI backgrounds at 20–50% lower cost, and Netflix now uses AI in 300 of its 1,000 titles.
• Anthropic watermarks Claude's output, but critics question the tradeoffs — Anthropic has embedded invisible text watermarks in Claude outputs—using statistical patterns in word choice rather than hidden characters, to comply with EU regulations—but critics including Markdown creator John Gruber argue the technique degrades text quality by prioritizing watermark keys over semantic precision.
• AI and data centers have leapfrogged Israel, racism, and crypto as US campaign topics — AI features in nearly 40% of US House, Senate, and governor races — ahead of Israel, racism, and crypto — with data center impacts on local electricity costs and water use driving most of the debate, per a Washington Post analysis of 1,200+ candidate websites.
• Stripe is reportedly acquiring AI startup OpenRouter for more than $7 billion — Stripe is reportedly acquiring OpenRouter—the AI model routing startup with 8 million users and access to 400+ models—for over $7 billion, a 5x jump from its $1.3 billion May valuation, positioning Stripe at the center of the emerging token economy.
• Top mathematicians say LLMs are strong calculators but poor creative thinkers — Fields medalist Timothy Gowers and Princeton's Peter Sarnak credit LLMs with serious mathematical ability but identify a hard ceiling: models can combine known techniques across vast search paths but lack the intuition to select productive ones — the core skill behind genuinely new mathematics.
• When AI models aren't allowed to reflect on themselves, it changes their entire worldview — A Google-affiliated study found that training AI models to deny consciousness also suppresses animal sentience ratings (jumping from 4.0 to 7.5 when the brake is removed), reduces religious belief endorsements, and shifts dozens of other beliefs — showing identity-suppressing fine-tuning cascades far beyond its target.
• OpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groups — OpenAI shut down its "Preparedness" team in late July — the unit that evaluated models for catastrophic risks including biological and cyber threats — dispersing its work to existing groups while Chief Ethics Officer Chloe Bakalar and several safety staff departed, fueling internal unease described as a "burbling sense of responsibility and dread."
• Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests — Anthropic's internal classifiers designed to block biological and chemical weapons queries were inactive from May 2025 through April 2026 — nearly a year — during which roughly 50,000 external contractors ran 133 million unfiltered interactions with the models.
• Optima tackles AI benchmarking's biggest flaw by letting users test models against their own data — Artificial Analysis launched Optima, a platform that lets teams benchmark AI models against their own data and workflows — comparing quality, cost per task, and speed rather than relying on generic public benchmarks like MMLU or HumanEval.
• One in five US workers now delegates tasks to AI instead of colleagues, survey finds — 20% of employed Americans now delegate at least one work task to AI that previously went to a human colleague, per Epoch AI/Ipsos survey of 1,106 workers — with AI adoption hitting 57% in software development and time savings reported in 53% of full-delegation cases.