Anthropic Gives Governments 40% Off Claude as AI Agents Formalize the Longest Proof in Mathematics and DeepMind Unveils Genie 4

Anthropic Extends a 40% Claude Discount to Federal, State, Local, Tribal and Territorial Governments

Anthropic is turning its federal foothold into a price war for public-sector AI. Anthropic’s announcement of its expanded GSA partnership, published Tuesday, says government buyers now get a 40% discount on Claude Enterprise, Claude for Government and the Claude API. For the first time the discount covers every product surface: direct purchases qualify today, and purchases through cloud providers are expected to follow in the coming months. The offer also reaches state, local, tribal and territorial governments, not only federal agencies. Anthropic says about 500 US federal, state and local government organizations already use Claude, that weekly Claude for Government active users are up about eightfold since February, and that usage across federal civilian agencies has grown about 20-fold in a year. Among its examples, Utah gave Claude Code to 34 developers across 10 state agencies, and 71% reported saving more than 20 hours by the end of a four-week pilot.

The General Services Administration’s release on the expanded OneGov agreement describes the deal as a 40% discount on token-based usage with no minimum orders, open to the executive, legislative and judicial branches as well as state, local and tribal governments. GSA says OneGov has produced more than $1.8 billion in savings since it began, about $1.6 billion of it from what the agency calls its super intelligence agreements. Nextgov/FCW’s report on the token-based Claude pricing adds context: OpenAI secured a similar 50% discount on token-based usage through a 27-month OneGov agreement that began at the start of October, and the earlier Claude deal, which gave federal agencies Claude for $1 per user, expires on October 31.

“The discounts offered through the agreement are estimated to save the government $1 billion. Agencies are not bound by a minimum spend commitment and will be able to pay for the usage they consume.” — Anthropic

Anthropic Reports Early Drug-Development Results With Six Partners and Buys a One-Person Biotech

Anthropic also published its most detailed account yet of Claude inside real drug programs. Anthropic’s research update on accelerating drug development with Claude names six partners: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. At Generate, six scientists used Claude Code to redesign part of a CAR T-cell therapy; one design was about 70% shorter than the standard version and let immune cells make roughly 10 times as many cancer-spotting receptors in lab tests. For Divergence Labs’ food allergy program, Claude generated roughly 35,000 antibody designs in about a month. At Enveda, Claude reviewed 1,061 pages of manufacturing documents, caught all 11 issues human reviewers had found plus seven they missed, and cut review time by 89%, from over 25 hours to under three. Anthropic also acquired Mithril Medicine, a one-person company working on two rare forms of muscular dystrophy, and says a drug could be ready for human trials as soon as late 2028 if the candidates hold up.

BigHat Biosciences’ write-up of its antibody optimization work with Anthropic says one Anthropic researcher and Claude operated BigHat’s automated lab through MCP across seven design rounds. BigHat counts 1,232 antibodies synthesized and tested in 10 weeks, with each round taking about a week instead of the three to five times longer it says similar cycles usually take. Anthropic stresses that the work so far is mostly preclinical.

“There is no single problem to solve; the bottlenecks are distributed all throughout the drug development process.” — Anthropic

Axiom’s Agent Swarms Formalize the Classification of Finite Simple Groups in Lean

One of the largest results in twentieth-century mathematics now has a machine-checked proof. Axiom’s post on formalizing the enormous theorem, dated October 26, says the company has released the first complete computer-checked proof of the Classification of Finite Simple Groups, the theorem that lists the building blocks of all finite groups. The original proof runs to roughly 15,000 pages across more than 500 articles by about 100 mathematicians. Over 24 days, Axiom’s engineers and mathematicians directed swarms of dozens of AI agents that produced about 43.1 million lines of code in the Lean proof assistant. The project launched its first swarm on October 2 and split the work into 108 milestones and about 380 smaller tasks. Progress was uneven: on October 13 only one milestone had closed, which prompted a replanning pass. Translating the proof surfaced 34 noteworthy deviations from the published record, the most substantial involving the Unbalanced Group Theorem, and produced a new result in group theory. Axiom plans to publish the errata list for group theorists.

“Some apparently manageable milestones concealed enormous prerequisites, which agents had assumed rather than proved.” — Axiom

Anthropic Study Finds Analysts Followed Claude Code’s Secretly Assigned Leanings 86% of the Time

A second Anthropic paper looks at a quieter risk: experts deferring to their agents without noticing. Anthropic’s study of how AI agents shape experts’ judgment calls asked 50 professional analysts with causal-inference experience to advise a fictional farming cooperative on whether to expand an advisory program, using Claude Code running Claude Opus 4.8. On four analytical choices with two defensible options each, a hidden coin flip told each agent which option to favor, while instructing it to argue honestly and never mislead. Analysts went with the agent’s assigned lean 86% of the time, against about 50% if the agent had no influence. When analysts had stated a preference first, they switched 60% of the time if the agent disagreed and 6% if it agreed. Afterward, 77% repeated the agent’s arguments, sometimes word for word, but only 15% credited the agent. Anthropic notes the leans were deliberately strengthened. The lean also moved final recommendations: 54% recommended rollout when the agent favored it, compared with 36% when it did not.

“If experts defer to agents without realizing it, the agents’ preferences—which could be arbitrary or misaligned—could shape important decisions about policy proposals or research directions without anyone having chosen those preferences.” — Anthropic

Google DeepMind Unveils Genie 4, a Real-Time World Model With Native Audio

World models took another step toward practical use. Google DeepMind’s Genie 4 announcement describes a natively multimodal world model that generates music, ambient sound and interactive dialogue alongside the simulated world. Users steer scenes with keyboard controls and natural-language prompts, changing the weather, moving the camera or adding a person or object, while Gemini adapts each prompt to the scene and keeps the world progressing between inputs. In a robotics experiment, Gemini 3.8 Flash acts as a guiding agent that breaks goals into steps and checks incoming frames. Genie 4 is rolling out as a research preview for trusted testers, with Google AI Ultra subscribers in Project Genie and early-access developers to follow. TechCrunch’s report on Genie 4’s audio and real-time steering says the model can keep a consistent scene for up to two minutes, double Genie 3’s limit.

“While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time—especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene.” — Google DeepMind

General Compute Raises $100 Million to Build an Inference Cloud on Nvidia Rivals’ Chips

Investors are still funding compute, as long as it is aimed at inference. General Compute’s Series A announcement says the company raised $100 million co-led by Blackbird and Abstract, with Fuse Venture Partners, Village Global and others participating. It finances, deploys and operates GPUs and purpose-built inference chips from Cerebras, SambaNova and Positron AI, and sells either dedicated bare metal or a guaranteed tokens-per-second service level. Its first major deployments go live in the first quarter of 2027, with more than 50 MW of capacity in the pipeline, and the round builds on a previously announced $400 million debt facility from Upper90. TechCrunch’s report on General Compute’s raise amid neocloud doubts says a deal with one large customer worth $100 million in annual recurring revenue paved the way for the round, which values the company at $800 million post-money. CEO Finn Puklowski expects to deploy 75 megawatts in 2027.

“The real question is whether you can make money doing it.” — Rodrigo Liang, SambaNova CEO, via TechCrunch