Live Feed
Ant Group Opens Ling 3.1 Flash Weights: 560B Hybrid-MoE Under MIT License With 1M Context
InclusionAI published full weights for Ling 3.1 Flash on October 14, two weeks after its free-trial API debut. The 560B model mixes 63 linear-attention layers with 9 gated MLA layers, targets 1M context, and runs on 8x H200 or 4x GB300 hardware under an MIT license.
Omnara Ships Self-Hostable Alternative to Claude Managed Agents and OpenAI Agents API
Omnara launches as an Apache 2.0 open-source managed agent platform that runs on any model and any machine, directly competing with Anthropic's and OpenAI's proprietary agent infrastructure.
OpenAI Releases First Official API CLI — Go-Built, Brew-Installable
OpenAI has published the first official command-line interface for its REST API on GitHub. Built in Go 1.26, installable via Homebrew, with resource-based command routing and an integrated image picker.
Mercury 3: Inception Scales Its Diffusion LLM 40x While Holding 232 Tokens per Second
Inception's Mercury 3 is more than 40 times larger than the original Mercury by parameter count and still generates at 232 tokens per second -- the same throughput as its predecessor. At $0.10 per million input tokens with a 1-million-token context window, it targets agentic pipelines where latency compounds across calls.
Thousands of GPT-Rosalind Agents Mapped Millions of Mouse Brain Cells in 32 Minutes
OpenAI, PerturbAI, and Scripps Research used distributed AI agents to process a whole-brain CRISPR atlas in 32 minutes, surfacing disease leads in Alzheimer's and Parkinson's that would have taken human researchers months to a year.
OpenAI Frontier Models Land on Google Cloud, Ending Azure-Only Era
OpenAI and Google Cloud are partnering to bring frontier models and Codex agentic workflows into Model Garden. Access extends to Google Antigravity and the Gemini Agent. Microsoft stays primary cloud partner.
GPT-5.6 Sol Found Critical PHP Zero-Days Before Attackers — OpenAI Commits $1B to Defenders
Cybersecurity firm Calif used GPT-5.6 Sol under OpenAI's Patch the Planet initiative to find critical PHP vulnerabilities before attackers. OpenAI is committing $1 billion in subsidized Daybreak access to frontline defenders.
Krea 3 Ultra Enters Arena Image Edit at #4 — $0.06 per 2K Image, 73% Below GPT Image 2
Krea's image editing model lands fourth on Arena's multi-image edit leaderboard at 1,435 ELO within days of launch. At $0.06 per 2K edit, it costs less than a third of GPT Image 2 at volume. The October 20 update adds an Ultra thinking mode to the API, FLUX 3 Max video up to 1080p, and an Agent API that can drive Blender.
Mistral Ships Voxtral 3: Voice Agent SDK with Sub-Second Turns Across 50 Languages
Mistral's Voxtral 3 combines transcription, reasoning, and speech synthesis into a single streaming pipeline completing each conversation turn in under one second. Five models cover batch and realtime STT plus TTS, with a Pro Voice mode that clones a speaker from 30 minutes of audio, starting at $0.0035 per minute.
Terminal X Built 1,700 Firm Investment Memories with GPT-6.1 Sol — 88.6% Match with Senior Human Reviewers
OpenAI case study: Terminal X encodes how private equity and credit firms make decisions as agents running on GPT-6.1 Sol. In 500 scored judgments, the Eval Agent agreed with senior reviewers 88.6% of the time — 220 basis points ahead of the next-best model, at 30% lower cost per query.
Anthropic: Claude Found 493 Attack Surface Issues Across 25 Open Source Codebases
New Anthropic security research scanned 5 million lines of code with Claude. Every one of 11 gated projects had code reachable before authentication that should not have been. 162 configuration combinations silently disabled security controls.
SlopCodeBench October 2026: Astra Leads at 16.3% as xHigh Effort Backfires for Opus 5.5 and Sol
Frontier models score 12-16% strict pass rates on SlopCodeBench's 36-problem evolving requirements suite. xHigh effort cut Opus 5.5 strict passes from 28 to 25 and cost 4.22x more. GPT-6.1 Sol Medium matches Astra Medium in strict passes at one-fourth the cost.
Claude Generated 318,484 Protein Designs in 64 Hours and Got 17 Lab-Confirmed Scorpion Antivenom Binders
Anthropic's research model autonomously designed proteins against scorpion neurotoxins that kill thousands annually. 17 of 211 tested binders passed experimental validation at Adaptyv Bio and Twist Bioscience. All designs are public.
Artificial Analysis Launches AA-Robotics v0.1: GPT-6 Astra Leads at 69.4 as Frontier LLMs Learn to Move Robot Arms
A new embodied AI benchmark from Artificial Analysis puts five frontier models in direct control of a 6-axis robot arm. GPT-6 Astra scores 69.4, Gemini 4 Argon 55.3, Claude Opus 5.5 48.8. Models are roughly 100x too slow and expensive for production use.
DHS Orders DeepSeek Removed from All Civilian Federal Systems, Backdated to July
The Department of Homeland Security published a Federal Register order Monday banning all DeepSeek AI from civilian executive branch agencies. The order covers any product built with DeepSeek models and is retroactive to the end of July. Defense agencies are excluded.