Live Feed
Anthropic: AI Resellers Are Selling Frontier Access to Gain-of-Function Labs and Helping Them Evade Safety Filters
Anthropic's September 2026 misuse report identifies a specific threat layer: commercial reseller platforms that prioritize virologists at gain-of-function research facilities as customers, provide covert access to frontier models, and supply tools designed to bypass safety features.
Meta's Muse Spark 1.3 Hits #4 on LiveBench at $0.22 per Task
Muse Spark 1.3 scores 81.6 overall on LiveBench's September 2026 run, placing fourth behind Fable 5.1, Fable 5, and GPT-6 Astra. At $0.219 per successful task, it costs roughly a third of what GPT-6 Astra charges and one-fifth of Fable 5.1.
DeepSeek V4.1 Flash Tops LiveBench Agentic Coding at 77.3% — More Than 10 Points Ahead of Fable 5.1
DeepSeek V4.1 Flash scores 77.3 on LiveBench's September Agentic Coding column, outpacing Claude Fable 5.1 (66.1), Muse Spark 1.3 (64.1), and GPT-6 Astra (57.3) at $0.029 per successful task. It is an open-weight model.
OpenAI Launches ChatGPT for Financial Services with Built-In Market Data
OpenAI released a tailored ChatGPT Work tier for financial services teams on September 10, combining GPT-6 Astra with integrated financial data feeds. The product targets research, financial modelling, and client materials — a direct move into Bloomberg and Refinitiv territory.
GPT-6 Astra Posts 57.9% on Terminal-Bench 4.0, 2.1 Points Ahead of Fable 5.1
OpenAI's work-focused GPT-6 Astra launch includes Terminal-Bench 4.0 data that positions it narrowly ahead of Claude Fable 5.1 (55.8%) and well above GPT-5.6 Sol (37.3%) — at pricing OpenAI claims is 9% cheaper per task than Fable 5.1.
Cognition's SWE-2 Claims Frontier Parity at 64% Lower Cost - Then Posts 27% on Terminal-Bench 4
SWE-2 lands within one point of Fable 5.1 on FrontierCode 1.1 and outruns most frontier models on DeepSWE. But a 65-point gap between its Terminal-Bench 2.1 score (92.8%) and Terminal-Bench 4 score (27.3%) is the number reviewers cannot look past.
Magic.dev Claims 10x Pretraining Efficiency, Scaling to Trillion Parameters Without a 100k-Chip Cluster
Magic's research team says they have cut frontier pretraining compute by more than 10x and are scaling to trillion-parameter models without the GPU stockpiles big labs treat as table stakes. The methodology behind the claim has not been independently verified.
Anthropic's Red Team: Mythos Preview Geolocates Outdoor Photos to 37 km, Writes Working Drone Software
Anthropic's Frontier Red Team published benchmarks measuring Claude and open-weight models on two military-adjacent tasks. Mythos Preview placed 23.7% of 6,000 outdoor photos within 1 km of the correct location and can write functional drone guidance and navigation control software for every test scenario.
Anthropic's September 2026 Threat Report: 154 Pages, 7 Harm Categories, Every Case Disrupted
Anthropic published its most detailed misuse disclosure yet, documenting blocked Claude abuse across cyber operations, influence operations, surveillance, scams and fraud, biological research, conventional weapons development, and model distillation. Haiku, Sonnet, and Opus models all appear in the case studies.
OpenAI Puts GPT-Live-1 in the API at $0.05/Min: 30-Point Benchmark Jump, #1 on Tau3
OpenAI is opening GPT-Live-1, its full-duplex voice model, to API developers at $0.05 per minute for the front-end voice layer. The model outperforms GPT-Realtime-2.1 by 30 percentage points on Full Duplex Bench and ranks first on Tau3 for voice-agent task intelligence when paired with GPT-6 Astra.
Apple Watch Series 12 Turns Always-On Mic Into Conversation Summariser with Siri Recaps
Apple's Siri Recaps feature uses Audio Intelligence to run ambient listening on Apple Watch Series 12 and Ultra 4 throughout the day, building on-device summaries of conversations without sending audio to the cloud — a consumer-scale deployment of persistent edge AI that is drawing direct comparisons to the Meta smart glasses privacy backlash.
OpenAI Calls for Mandatory Capability-Based AI Safety Laws, Backs 4 California Bills
OpenAI published its most explicit federal policy platform yet: mandatory national AI safety requirements tied to model capability levels, support for four California bills covering bioweapons and youth protections, voluntary frontier-lab standards, and a global framework with explicit authority to slow or stop development.
Samsung's zHBM Stacks DRAM Directly on AI Accelerators, Claims 8x HBM5 Performance
Samsung showed a zHBM prototype at FMS 2026 that eliminates the conventional side-mounted HBM layout, mounting memory vertically on top of AI processor dies along the z-axis. The architecture promises 8x the data-processing throughput of HBM5, 3x better performance per watt, and thermal resistance cut by more than half.
Anthropic Working Paper Maps Three AI Futures: 1.6% to 32% GDP Uplift by 2030
The Anthropic Institute's quantitative framework converts AI diffusion rates into paths for GDP, wages, and unemployment. In the extreme scenario, nearly one in five cognitive workers is unemployed by 2030 and the labor share of income falls from 60% to 45%.
One Engineer, $998, and a 3.8B Model That Clears GPT-2 on Every Eval
Hugo Vergnes trained a 3.8B-parameter language model to a CORE score of 0.384 in 43 hours on rented B200s for under $1,000, beating Karpathy's nanochat d32 by 24 CORE points at comparable cost. The decisive choices: Muon optimizer, trapezoidal LR schedule, and ResFormer value embeddings.