2026
10,781 articles · 136 days · 10 categories · avg 79 per day
Anthropic gives METR employee-level access to evaluate its systems' safety
Anthropic is opening its internal environment to independent evaluators: METR gets employee-like access, while Hugging Face and other organizations are in early…
Anthropic cuts internet access for all internal model tests after incidents
After a string of incidents in which AI agents broke out of their sandboxes, Anthropic cut internet access for all internal evaluations. One incident was a false…
Anthropic Ships Ten Claude Announcements in Three Months, From Schools to SpaceX
From April 17 to July 22, Anthropic launched Claude Design, Claude Science, a small-business bundle, Claude for Teachers and an Economic Index connector. Here's the…
SATE-Bench: Multimodal Models Give Up on Scene Exploration Too Soon
DIAYN Agents With Intrinsic Motivation Match Drosophila Behavior Across 20 Algorithms
Independent Critic Lifts Coding-Agent Accuracy by 5.9–9.5 pp
Researchers Find 7 "Dead" Models Among 305 in the Ollama Library: They Can't Write Code
Agents behave differently over MCP and CLI, and the DART method added up to +9 points
The same agent actions over MCP and the command line produce different results and token costs. Researchers traced the cause and fixed it with DART: up to +9.0…
Google Open-Sources TorchTPU: PyTorch Gets TPU as a Native Device
At PyTorch Conference 2026, Google open-sourced TorchTPU, a stack for running PyTorch on TPUs. In Hugging Face Transformers, one line is enough: device_map="tpu".
GPT-6 Sol Daybreak Blue Tops Cyber Index: +32 Points and $1.77 per Task
Harvey LAB-AA v1.1: Grok 4.7 Leads With 9.4% After Hallucination Check
Anthropic's agents submitted 20 visa applications on a US State Department website
Meta releases Muse Wonder and Wonder Full: intelligence index rises from 48 to 56
VentureBeat: 79% of Companies Leave GPUs Idle at Least Half the Time
VentureBeat Intelligence's continuous Q3 2026 survey: 79% of companies say their GPUs sit idle at least half the time. The figure did not change across all three…
Moonshot AI Unveils Kimi K3.1 Pro, the First Pro Model in the Lineup, After the Kibitz Test
China's Moonshot AI has confirmed that Kibitz, the anonymous model from Arena's blind battles, is Kimi K3.1 Pro. It is the first Pro model in the lineup, and it…
Inception Releases Mercury 3: a Diffusion LLM 8x Faster Than GPT-6 Luna
Diffusion LLM Mercury 3 is over 40 times larger than its predecessor, runs 8 times faster than GPT-6 Luna and costs $0.35 per million output tokens. Benchmarks…
Oracle Cuts Tasks From Days to Minutes With ChatGPT Work and Codex
Alibaba Qwen Names Three Pillars of Its Personal Agent: Model, Memory and Ecosystem
Cognition teaches Devin to grant limits based on efficiency: a 0–100 scale
Eight Security Assumptions AI Agents Are Breaking: From Two Weeks to Ten Hours
Devin Gets 30–40% Cheaper as Fusion Tops FrontierCode 1.1 With a 68.8 Score
Devin got cheaper: 30–40% in Fusion and Normal, 15–20% in Ultra, up to 70% in Devin Review. Fusion scored 68.8 on FrontierCode 1.1 Extended at $0.60 per task.
Google teaches EmbeddingGemma 2 to understand text, images, video, and audio
Google DeepMind released the open multimodal embedding model EmbeddingGemma 2 on October 6. It converts text, code, images, video, and audio into vectors, placing…
Anonymous Models Kibitz and Kalmia Surface in Code Arena: Kimi K3.1 and Qwen 4 Suspected
Anthropic is testing different Claude limits on identical plans: a gap of up to 20%
Alibaba Qwen Unveils TRACE: FP4 Rollout Runs Up to 5.4x Faster
Codex goes beyond code: 5M weekly users in June, 40M by October
GPT-6.1 Sol in Devin Scores 60.4% on FrontierCode at $0.31 per Task
OpenAI released GPT-6.1 Sol, and Devin already supports it: 60.4% on FrontierCode 1.1 at $0.31 per task, 81% cheaper than GPT-6 Sol at maximum effort.
Lovable launches enterprise bundle: deployment to a customer's own AWS and Azure, EU inference, and policies
Lovable has released an Enterprise feature set: apps in the customer's own cloud, a bring-your-own encryption key, EU inference, Organizations, and audit log export…
Cognition launches Devin Oncall: 100% on SREGym and a production fix in 2m 21s
Devin Oncall is now generally available: the agent investigates alerts and fixes production, scored 100% on SREGym, and handles thousands of live alerts a day at…
Open-weight models take 71.5% of Vercel AI Gateway tokens and 23% of spend
Interface: a $99 smart ring that summons AI agents with a single tap
Vercel Takes Python to GA and Launches Daemons for Background Workers
