#benchmarks

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.

OpenAI quietly launched GPT-5.6 in three tiers (Sol, Terra, Luna) restricted to trusted partners, while Semgrep reported that GLM 5.2 outperforms Claude on its internal cybersecurity benchmarks. Portugal's National Health Service contracted Sword Health to deliver AI-assisted physical therapy nationwide, the first country-scale deployment of its kind. Edison Scientific and Population Health Partners announced a deal to use AI agents for end-to-end drug discovery. Patrick Boyle assessed Brexit's 10-year economic toll, and Noah Smith examined whether AI will push companies toward more or less outsourcing.

Anthropic accused Alibaba of illicitly extracting Claude's model capabilities, drawing 829 Hacker News comments and raising fresh questions about AI IP enforcement. OpenAI unveiled Jalapeño, its first custom inference chip built with Broadcom, a move that reshapes its dependency on Nvidia. GLM-5.2 crossed a threshold Nathan Lambert called the step change for open agents. Google introduced computer use in Gemini 3.5 Flash, and U.S. health spending rose 7.3% in 2025, driven by GLP-1 drug adoption and higher care utilization. A bipartisan bill would expand methadone access by letting physicians prescribe it outside specialty clinics.

AbbVie agreed to acquire Apogee Therapeutics for nearly $11 billion, the latest in a pharma M&A wave that has totaled $123 billion in 2026 deals so far. Definium's LSD therapy met its primary endpoint in a Phase 3 trial for major depression, and the FDA reversed a four-month-old rejection of Regenxbio's gene therapy for Hunter syndrome. In AI, Samsung deployed ChatGPT Enterprise and Codex to employees worldwide in one of OpenAI's largest enterprise rollouts, while Nathan Lambert and Kevin Xu argued that banning open-source AI would damage US interests. Patrick Boyle examined how SpaceX humiliated Wall Street and Cuba announced its most sweeping economic reforms in decades, including private banking and the end of fuel subsidies.

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.

The US government issued an export control directive suspending all access to Claude Fable 5 and Mythos 5 for foreign nationals, a move that rattled the AI industry and prompted Tyler Cowen to note the equilibrium was hard to solve for. Separately, Arvind Narayanan and Sayash Kappoor published a detailed argument that software engineers won't be replaced by AI, citing task fragmentation and the compounding cost of errors. The FDA approved Sanofi's teplizumab for children with stage 3 type 1 diabetes after it cleared a dispute between career staff and political appointees. Noah Smith asked whether Chinese corporate debt could zombify the economy, while Patrick Boyle examined how SpaceX humiliated Wall Street by retaining founder control across 25x more fundraising than typical startups.

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.

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.

Sebastian Raschka published a comprehensive 2025 LLM review covering DeepSeek R1, RLVR, inference-time scaling, and 2026 predictions, while Simon Willison released his annual year-in-LLMs retrospective. On the pharma front, NIH Director Jay Bhattacharya confirmed that DEI-related grants restored under a court order will not be renewed in 2026, even as NIH began reviewing thousands of delayed proposals and funded 135 on the first day. Novo Nordisk priced its newly approved Wegovy pill below Eli Lilly's competing obesity drug in a direct-to-consumer push. On the economy side, researchers across 16 countries found that the post-1980 labor share decline is driven by the composition of capital, not its quantity, and building more structures may reverse it.