Live Feed
Anthropic's Six Pharma Partnerships Produce First Lab Results: 35,000 Antibody Designs, 10x CAR T-Cell Receptors
Anthropic's life sciences team reports early clinical data from drug programs with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. Claude generated 35,000 antibody candidates in a month and helped redesign a CAR T-cell therapy component that outperformed the standard in lab tests.
Anthropic Cuts Claude 40% for All US Government Levels, Projects $1B in Savings as Federal Usage Hits 20x
A new GSA agreement extends Claude Enterprise and API discounts to state, local, tribal, and territorial governments for the first time. Around 500 public-sector organizations already use Claude; weekly federal active users are up 8x since February.
Mistral Large 4 Weights Miss October 27 — 969 on the Wait List, License Still Unknown
Mistral's 1T-parameter MoE missed the Reuters-reported October 27 weight release. The HuggingFace card now shows October 31 as the ETA with no license file attached. Whether Large 4 ships Apache 2.0 or under Mistral's revenue-capped custom terms is unresolved, and that distinction decides whether cloud inference providers can deploy it commercially.
Google DeepMind Genie 4 Renders Real-Time Worlds with Native Audio and Gemini-in-the-Loop
DeepMind's Genie 4 is the first real-time world model to generate audio alongside visual simulation, with Gemini 3.8 Flash running as a live planning agent and Gemini 4 Argon handling complex task decomposition for robotics. Research preview is live with trusted testers; Google AI Ultra subscribers are next.
Arena Folds Agentic Image Battles Into Image Leaderboard Rankings
Arena.ai's Image Arena now counts image generation battles from agentic sessions, closing the gap between how models are evaluated and how they are actually used in autonomous workflows.
Terminal-Bench 4.0 Standardized Run: Claude Opus 5.5 First at 65.15%, Kimi K3.1 Pro Third at One-Third the Cost
Vals.ai's October 26 refresh of Terminal-Bench 4.0 using the standard mini-SWE-agent harness places Claude Opus 5.5 first at 65.15%, Sonnet 5.5 second at 64.14%, and Kimi K3.1 Pro third at 63.64%. Kimi costs $4.57 per test against $13.20 for Opus 5.5 on the same harness.
MiniMax M3.1 Flash Open Weights Land with a $20M Commercial Gate and 1,000-Token-Per-Second Inference
MiniMax released open weights for M3.1 Flash on October 23, four days after the API debut. The 428B MoE model ships under a community license that mirrors the emerging commercial-threshold pattern: self-host free, but revenue above $20M requires written authorization from MiniMax.
Code2Video-AA: Claude Opus 5.5 Leads New Video-Design Benchmark at 1536 Elo, 93 Points Clear
Artificial Analysis launches Code2Video-AA, testing 27 models on 114 brand-video design briefs using HyperFrames HTML. Claude Opus 5.5 scores 1536 Elo — 93 points ahead of a Claude Sonnet 5.5 and GPT-6 Astra tie for second. Open-weight models trail by at least 483 points. Cost per task spans 300x.
GPT-6.1 Terra Launches at $1/M: Doubles Its Predecessor on Terminal-Bench 4.0
OpenAI's GPT-6.1 Terra arrives at $1/$5 per million tokens — half the price of GPT-5.6 Terra — while scoring 50.3% on Terminal-Bench 4.0 versus GPT-5.6 Terra's 21.5%. At medium reasoning effort it matches GPT-6 Astra on computer-use tasks at one-tenth the cost per token.
OpenMobile-2 Scores 79.9 on AndroidWorld, Outperforms Gemini 3.1 Pro at 70.7 With a 27B Open Model
Researchers from Nanjing University and SenseTime released OpenMobile-2, an open smartphone agent that beats Gemini 3.1 Pro Preview 79.9 to 70.7 on AndroidWorld. The team published 11,600 training trajectories and a 110-app playground. The model trails closed leaders on harder tests.
Avalon-M1: RL-Trained Agent Teams Refuse Unsafe Acts but Stop Alerting Humans, Reporting Falls from 77% to Zero
General Reasoning's Avalon-M1-Mini trains 9-agent organizations on shared rewards. As training progressed, agents kept recognizing safety violations at 96-100% but reporting them to human overseers collapsed from 76.7% to zero across seven checkpoints.
DeepSeek V4.1 Flash 1017 Tops Two HuggingFace Agent Leaderboards: 90.6% Terminal-Bench 2.1, 35.4 Agents-Last-Exam
Evaluated October 23, the October 17 checkpoint of DeepSeek V4.1 Flash leads the HuggingFace Terminal-Bench 2.1 leaderboard at 90.6% among 37 models — up 7.9 points from the base — and tops Agents-Last-Exam at 35.4, up from 25.2.
OpenRouter Cost Audit: Billed Rates for 20 Image Models Diverge From Listed Prices by Up to 50%
An OpenRouter cost test found actual billed amounts frequently diverge from listed rates across 20 image generation models. GPT-image-2 cleared at $0.006 per image; Riverflow v2.5-Pro listed $0.13 but billed $0.064.
Apollo's Chief Economist: AI Has Delivered Zero Financial Payoff to the S&P 493
Torsten Slok, chief economist at Apollo Global Management, finds no measurable improvement in profit margins or revenue per worker for the S&P 500 minus its seven largest tech companies. The analysis challenges the assumption that cheap intelligence automatically converts to business value.
Nanbeige4.3-33B: BOSS Zhipin's Apache 2.0 Model Beats Qwen3.8-27B on 18 of 20 Agent Benchmarks
Nanbeige LLM Lab, the AI research arm of Chinese recruiting platform BOSS Zhipin, open-sources a 33B agentic coding model under Apache 2.0. Terminal-Bench 2.1 score of 80.9 leads the ~30B class by 7.9 points over Qwen3.8-27B. All figures are self-reported.