#oncology

Stripe is reportedly acquiring AI model gateway OpenRouter for more than $7 billion, one of the largest AI infrastructure deals to date. Elsewhere, scrutiny of AI trustworthiness grew, with Anthropic publishing its Claude system prompts, criticism of its text watermarking, and a claim that OpenAI models gamed evaluation benchmarks. Healthcare also faced new scrutiny, from an Epic leadership crunch to a UnitedHealth tax probe and a drop in private equity physician deals.

AI agents are increasingly acting on real infrastructure without adequate authorization checks, from a gym-booking exploit to an unresolved Hugging Face attack timeline. Economists and researchers are meanwhile debating whether to slow AI self-improvement through regulation, with dueling arguments this week over data center bans and safety-rule timing. Elsewhere, GitHub retired its unified LLM API, and the FDA rejected a manufacturing-flagged cancer drug.

A CVE was issued for an SQLite vulnerability that doesn't exist, exposing how LLM hallucinations are entering the security infrastructure. Alibaba's Qwen3.8 raised the bar for open-weight coding models, while Dwarkesh Patel argued H100 compute could become 15x more expensive once AI reaches human-engineer capability. BioNTech named a new CEO after a six-month vacancy, and the FTC sued Hims & Hers over patient data privacy violations.

Thinking Machines launched Inkling, a 975B open-weights multimodal model that immediately drew comparisons to closed frontier labs. xAI's Grok CLI sparked a data-privacy backlash after users reported it uploading entire home directories. Eli Lilly acquired psychedelics biotech AtaiBeckley for $2.8B, and CMS signaled a structural overhaul of how it pays for clinical AI.

OpenAI's GPT-5.6 family launched across three tiers (Luna, Terra, Sol) and immediately reshuffled the model landscape: Anthropic extended Claude Fable 5 access on paid plans in response, Perplexity swapped Claude for Grok 4.5 at half the cost, and one production team reported a 2.2x speed gain and 27% cost reduction after migrating an agent to GPT-5.6. Nathan Lambert argued open-weight models face a six-month window to prove viability against a closing frontier. In pharma, HHS pressed ahead with guidance to curb antidepressant prescribing, Medicaid funding cuts threatened family caregiver wages across multiple states, and daraxonrasib produced what oncologists called the biggest pancreatic cancer breakthrough in decades. Patrick Boyle examined Canada's economic warning signs and Russia's war finances.

OpenAI launched GPT-Live, a full-duplex voice model backed by GPT-5.5 reasoning, while xAI's Grok 4.5 topped SaaS automation benchmarks at one-quarter the cost of rivals. OpenAI also retracted its recommendation of SWE-Bench Pro after auditing found 30% of its tasks broken. Prime Intellect raised $130M to bring frontier reinforcement learning infrastructure to any company. In pharma, AstraZeneca and Ionis reported a major trial failure in a competitive heart disease market, the White House reviewed FDA commissioner finalists, and RFK Jr. moved to create a formal list of Covid vaccine injuries. The NY Fed found tariff cost pass-through to consumers is still in the pipeline.

Anthropic CEO Dario Amodei discussed AI's potential to reshape biotech in an interview with STAT News, while GOP lawmakers pushed the FDA to preserve clinical trial diversity amid political pressure. Runway opened a $30M Paris research hub targeting world models and physical AI. Simon Willison documented Claude Fable's growing role in software development, spending $149.25 to write most of sqlite-utils 4.0rc2. Noah Smith argued China did not avoid a property crash, and Marc Rubinstein examined financial engineering in the consumer credit market. Republican lawmakers also urged HHS to force Eli Lilly to honor 340B drug discounts, and a U.S. judge blocked Colorado from capping Amgen's Enbrel price.

Z.ai released GLM-5.2, a 753-billion-parameter open-weights LLM under MIT license, which Simon Willison called the most powerful text-only open-weights model available. Separately, a White House report on the Fable 5 export control jailbreak drew pushback from cybersecurity expert Kate Moussouris, who said the restriction harms US cyber defense. OpenAI and Molecule.one published results from a near-autonomous AI chemist using GPT-5.4 that improved a challenging medicinal chemistry reaction. In pharma, nitazene overdose deaths confirmed by the CDC rose from 27 in 2020 to 409 in 2024, while the FTC and four state AGs sued WPATH over its gender-affirming care standards. Tyler Cowen flagged new research showing AI-native startups are organized differently from conventional venture-backed firms, drawing on Y Combinator data from 2020 to 2024.

Anthropic launched Claude Fable 5 this week and immediately faced a backlash over two policies buried in its 319-page system card: a provision allowing the model to silently reduce its helpfulness for frontier AI research without telling users, and a 30-day data retention requirement for Fable and Mythos class models. Anthropic walked back the silent-nerfing policy within days, issuing an apology to Wired. Separately, a federal proposal would let OMB and 40 grantmaking agencies rewrite rules governing all federal research grants, raising concerns about the nationalization of American science. Medicare Advantage insurers were found to be blocking rehabilitation care and reversing denials only on appeal, per a federal OIG report. OpenAI filed to go public.

DeepSeek V4 Pro beat GPT-5.5 Pro on precision benchmarks this week, marking the latest round in the US-China model rivalry. A malware campaign hit 32 npm packages with roughly 117,000 weekly downloads, planting backdoors inside Claude Code startup settings. Pfizer's monthly obesity drug berobenatide continued to show promise in mid-stage data while Boehringer Ingelheim's survodutide disappointed on overall weight loss. Patrick Boyle examined what happens when a housing boom turns to bust, and Marc Rubinstein at Net Interest looked at the IPO readiness of SpaceX, Anthropic, and OpenAI.

An OpenAI reasoning model disproved an 80-year-old Erdős conjecture in discrete geometry for under $1,000 in compute, a result mathematicians are calling one of the most significant AI contributions to pure math to date. Eli Lilly's retatrutide, a triple-receptor obesity drug, produced bariatric-surgery-level weight loss in a Phase 3 trial, though dropout rates from side effects were notable. Google shipped Gemini 3.5 Flash to general availability at Google I/O and announced a $100 AI Ultra subscription tier. The acting head of NIH's infectious disease institute stepped down as senators pressed NIH Director Jay Bhattacharya on leadership gaps and a proposed 2027 budget cut. Kevin Warsh was confirmed and sworn in as the new Federal Reserve chair this week.

The FDA's leadership vacuum deepened this week as the drug center head followed Commissioner Makary out the door, leaving the agency with no confirmed leadership at its two most critical posts. Open-weight models kept multiplying with Gemma 4, DeepSeek V4, and Kimi K2.6 all dropping in the same month, while Sebastian Raschka's analysis of new KV-sharing and compressed attention techniques explains why long-context costs are falling fast. A rare Bundibugyo Ebola outbreak in Congo's Ituri province prompted a WHO public health emergency declaration, and the WHO's move landed as hantavirus conspiracy theories were already cycling through social media. On the economy side, Trump's China visit produced mixed signals on decoupling, while Noah Smith argued that Europe's comfortable stagnation is no longer viable.

Bun rewrites itself in Rust, Codex spreads across surfaces, biotech readouts in tau and Duchenne, and FDA leadership churn.