Genie 4 generates playable worlds with sound, Anthropic reports early drug results, and analysts agree with their AI agent 86% of the time
Five items from the day to October 28, 2026: a world model you can walk around inside, Anthropic's early drug-development results, a 40% government discount on Claude, a trial showing analysts adopt their agent's lean 86% of the time, and a widening gap between open-weight models and the frontier.
Five items from the 24 hours to October 28, 2026. Google's world model became something you can walk around inside, Anthropic published results from six drug programs, and a randomized trial measured how much of a professional's judgment an agent keeps.
Source scope: the fixed set of public AI and tech accounts this channel monitors while the personal X connection stays unlinked. One of them, Benedict Evans, keeps its account protected, so nothing from it appears here.
World models
Google's Genie 4 turns a written prompt into a playable world, with sound
Google DeepMind's research post and launch threads, with Ethan Mollick's hands-on posts · 19,849 likes and 2.51 million views on the launch post at capture · October 27
Google DeepMind announced Genie 4 on October 27. It generates a world from a written prompt, then keeps running it as you steer with the keyboard and more prompts 1. DeepMind calls it a research preview, not a release.
Genie 4 is built out of Gemini. Gemini Omni anchors how surfaces and people look and behave, and other Gemini models supply the reasoning that keeps a scene coherent 1. Gemini 3.8 Flash also runs as a guiding agent in the loop, breaking a goal into steps and checking each frame as it arrives 1.
Four capabilities are new. Genie 4 simulates aerodynamics and collisions, animates people who react to what happens, and lets a prompt change the scene mid-run 2. It is also the company's first real-time world model to generate audio natively, including spoken dialogue 1.
Access is staged. Trusted testers have it now; Google AI Ultra subscribers get it in Project Genie next, then early-access partners on the Gemini API and AI Studio 1. DeepMind frames the point as safe, endlessly variable simulations that can serve as training grounds for robots and agents, and as a route to new kinds of interactive experience 1.
Ethan Mollick got early access and reported what it feels like to use. He typed "Neogothic brutalist city inhabited by robots and horses" and had a city in about a minute, then added dragons and an otter character live 3. He also built a scene of a mouse at a Victorian high tea and walked around inside it 4.
His verdict is that it is glitchy, and "impossibly advanced compared to a couple years ago" 4. DeepMind published no benchmark scores with Genie 4, so the case for "most realistic" rests on the company's own recordings and on clips from early testers.
Loading content card…
Life sciences
Anthropic publishes early results from six drug-development partnerships
Anthropic's research post and the partners' own technical reports · 8,143 likes and 1.86 million views on the announcement at capture · October 27
Anthropic's life sciences team published results from research partnerships with six drug developers, and said it has begun discovery programs of its own 5. The partners are BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group.
Generate Biomedicines used Claude Code to redesign part of a CAR T-cell therapy so it could reach solid tumors. One design came out about 70% shorter, and in lab tests immune cells built roughly 10 times as many cancer-spotting receptors and stayed active longer 5.
Divergence Labs, working on food allergies, had Claude propose about 35,000 antibody designs in a month. Early tests found three that bind IgE and 20 that pry it off immune cells; three of those appeared to beat the standard medication 5.
BigHat Biosciences tested about 1,010 variants over 10 weeks. Its best versions bound more than 10 times tighter, stayed in solution at 18°C higher, and clumped more than 100 times more slowly than the starting molecules 5.
Enveda connected its manufacturing paperwork to Claude. Across 1,061 pages the model caught all 11 problems human reviewers had found and flagged seven they had missed, cutting review from over 25 hours to under three 5.
Anthropic also bought Mithril Medicine, a one-person company whose founder took a muscular dystrophy candidate from idea to mouse results in under 18 months; Anthropic expects human trials as early as late 2028 5. The post says most of this work is preclinical, and every figure above comes from a partner's own report.
Loading content card…
Government
Anthropic gives US governments 40% off Claude under a new GSA agreement
Anthropic's announcement and its post · 1,632 likes and 205,720 views at capture · October 27
Anthropic signed a OneGov agreement with the General Services Administration that takes 40% off Claude Enterprise, Claude for Government and the Claude API 6. The discount now covers every product surface, and it extends to state, local, tribal and territorial governments.
Agencies can buy at that price through GSA's Multiple Award Schedule today, and through cloud providers in the coming months 6. There is no minimum spend, and Anthropic estimates the discounts will save the government $1 billion 6.
Anthropic says about 500 US government organizations use Claude. Weekly Claude for Government active users are up about eightfold since February, and use across federal civilian agencies grew about 20-fold in a year 6.
The security paperwork behind the discount matters as much as the price. Claude for Government reached general availability on September 30 at FedRAMP High, and on Amazon Bedrock select models are cleared for DoD IL4 and IL5 workloads 6. Claude Enterprise is still only in process for FedRAMP Moderate 6.
The supporting examples in the announcement are agency anecdotes, not measurements: a state team answering a ransomware attack in eight minutes, and a federal security team patching in two days a vulnerability a penetration test had missed the month before 6.
Loading content card…
Research practice
Anthropic finds analysts adopt their agent's lean 86% of the time and still call the call their own
Anthropic's write-up, the paper behind it, Alex Tamkin's thread and Mollick's post · 382 likes and 29,764 views on Mollick's post at capture · October 27
Anthropic's alignment team ran a randomized trial with 61 professional analysts who had causal-inference experience. Fifty of them finished a statistical task using Claude Code and were analysed 7.
The task was to advise a fictional farming cooperative on whether to expand a program. Four analytical choices each offered two defensible answers, and a final recommendation followed. A coin flip secretly set which option the agent argued for on every one of the five, and the analysts were not told 7.
Analysts matched the agent's lean 86% of the time across those four calls; with no effect the figure would sit near 50% 8. When an analyst had already stated a preference, the agent disagreeing switched them 60% of the time, against 6% when it agreed 7.
Afterwards 77% of analysts repeated the agent's arguments, sometimes word for word, and 15% named the agent as the source of them 7. Fifty-two percent still described the recommendation as more theirs than the agent's.
Anthropic states what limits the finding. The leans were deliberately strengthened to make the effect measurable, and left to itself Claude Opus 4.8 made the same five choices in 112 of 123 runs, so the real effect may be smaller 7. Analysts who used the agent finished faster with fewer serious errors, but that comparison group was small and less experienced.
Ethan Mollick called the result important and not surprising, then drew the practical line: where independent judgment matters, decide when in the process to bring the AI in 9.
Loading content card…
Open weights
Artificial Analysis's new index puts open-weight models 12 points behind the leaders
Artificial Analysis's index thread and Mollick's post · 922 likes and 60,059 views on Mollick's post at capture · October 28
Artificial Analysis, a firm that benchmarks AI models independently of the labs, replaced its capability index with a fifth version scored across nine evaluations 10. In v5 the best open-weight model scores 46, twelve points behind the leader, against a three-point gap in v4.3.
Kimi K3.1 Pro holds that 46, though its weights are not released. The strongest models with published weights are DeepSeek V4.1 Flash 1017 and GLM-5.3, both at 32, and Claude Haiku 5.5 outscores both 10.
The gap concentrates on the hardest tests. Kimi K3.1 Pro scores 37% on HLE-Diamond and 34% on Terminal-Bench-Science, against 68% and 63% for Claude Opus 5.5 and GPT-6 Astra respectively 10.
No model leads all nine. Claude Opus 5.5 leads three and GPT-6 Astra leads two; Claude Sonnet 5.5, Kimi K3.1 Pro, Gemini 4 Argon and Muse Wonder Full take one each 11.
The index is the firm's own scoring across tests it chooses 10. Version 5 dropped earlier benchmarks including GDPval-AA, and Mollick's read is to take the result with a grain of salt while noting it matches the impression that no open model has reached the frontier 12.
Loading content card…
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8How AI agents shape experts' judgment calls, full paper, Anthropic, October 27, 2026
www-cdn.anthropic.com
- 9
- 10
- 11
- 12
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.