Live Feed
Xiaomi MiMo V2.6 Pro and Flash Enter Agent Arena, Open-Weight Models Join the Agentic Leaderboard
MiMo V2.6 Pro (1.02T MoE, 42B active, MIT) and Flash (309B MoE, 15B active, MIT) were added to Arena's Agent leaderboard on October 15 — putting Xiaomi's open-weight models into direct head-to-head agentic evaluation against Claude, GPT-6, and Gemini.
Black Forest Labs Ships FLUX 3 Image Klein: A 5.6B Open-Weight Model That Runs on Gaming PCs
Black Forest Labs released the weights for FLUX 3 Image Klein on October 15, four days after quietly pushing them to Hugging Face. The 5.6B model supports both generation and editing, runs inside an 11.4 GiB heap on consumer hardware, and is priced 5x below the closed FLUX 3 Image API.
Anthropic Carries a BB+ Junk Rating as It Arranges $60 Billion in New Financing
Kroll Bond Rating Agency assigned Anthropic a BB+ issuer rating in June — the top of speculative grade, one notch below investment — and kept the outlook stable. The rating was never disclosed publicly until this week, as the company moves toward a $60 billion financing round.
Claude Code Review Exits Research Preview with 6-Minute Review Times and a 200% Code Output Gain
Anthropic made its agentic PR reviewer generally available on October 15, eight months after its research preview launch. The system dispatches a team of Claude agents per pull request, dropped median review time from 20 minutes to 6-10, and is credited with tripling daily code commits in Anthropic's monorepo.
Amazon in $7B Talks to Acquire Decart, Betting Its Optimization Stack Can Break CUDA Lock-In on Trainium
Amazon is in advanced talks to pay roughly $7 billion for Decart, a 100-person startup whose real value isn't its generative video products but software that lets AI models run on Trainium chips at half the GPU cost without a rewrite.
Postgres 19 Slips to November After AI Code Review Generates 40-Bug Backlog Human Maintainers Cannot Clear
The PostgreSQL Global Development Group pushed its next major release to November 19 and dropped 8-10 features after LLM-assisted review surfaced bug queues faster than volunteers can fix them — exposing the throughput gap between AI bug-finders and human fixers.
NVIDIA Nemotron 3.5 Bolt 74B Fits in 24GB at 2 Bits: Tops Its Class on GPQA Diamond, BrowseComp, Terminal-Bench
NVIDIA releases Nemotron 3.5 Bolt 74B-A9B on October 15, a Mamba-2-Transformer hybrid MoE compressed to 24GB via 2-bit quantization. It scores 86.4 on GPQA Diamond, 66.8 on BrowseComp, and 51.0 on Terminal-Bench 2.1 — above Qwen3.6-35B-A3B and Gemma4-26B-A4B on all three.
Third Moment Research Runs 10,000 AI Agents for 36 Hours Straight Hunting Quant Signals
The quantitative research firm deploys GPT-6.1 Sol as its primary workhorse for token efficiency, GPT-6 Astra for peak performance tasks, and GPT-5.6 Terra for targeted security audits. Co-founder Jeffrey Yan says Sol matches Astra output while using the fewest tokens.
One Developer Used AI to Clone Photoshop, Premiere, and Now Microsoft Office
Brandon Thomas shipped working open-source replacements for Adobe's Creative Suite using AI-assisted code generation. He is now releasing free equivalents of Word, PowerPoint, and Excel. The question is whether AI vibecoding has finally made proprietary software economically obsolete to replicate.
Grok Bot Adds Cross-Agent Memory: Bots Recall Context Across Quarters and Hand Off Projects Without Prompting
xAI's Grok Bot now demonstrates agents that pull multi-quarter context without user input and transfer running projects between bots autonomously. The $30/month product learns workflows by demonstration and runs them on schedule.
Claude Opus 5.5 Claims #1 on Agent Arena as Muse Wonder and Haiku 5.5 Make Their Debuts
Anthropic's Opus 5.5 Max leads Arena's 54-model Agent Arena at 17.2 percent net improvement after jumping 16 spots. Meta's Muse Wonder enters at #12 with the lowest tool hallucination rate of any new entrant this week. Claude Haiku 5.5 enters at #27 at $0.10 per completed task with 21,839 sessions already on the board.
Muse Wonder Full Scores 56 on AA Intelligence Index, Ranks Second Behind Claude Opus 5.5
Three days after a benchmark-free launch, Muse Wonder Full gets its first independent score: 56 on the Artificial Analysis Intelligence Index v4.3, second globally behind Claude Opus 5.5 at 58 and ahead of Claude Fable 5.1 and GPT-6 Astra. Meta's 6-week jump from Muse Spark 1.3 at 48 is 8 index points.
GLM-5.3 Reaches 95.4% on SWE-bench Verified in Independent vals.ai Testing
vals.ai's neutral bash-only harness puts Z.ai's GLM-5.3 at 95.4% on SWE-bench Verified, placing it inside the top six globally alongside Claude Opus 5, DeepSeek V4 Pro, and GPT-5.6 Sol. It is the second Chinese lab model above 95% on an independent benchmark.
DeepSeek V4 Pro's Week-Long $0.21 Rate Was a Data Error, Not a Price Cut
Five consecutive observations of the DeAI Price Index confirm last week's $0.2088 input / $0.4176 output reading for DeepSeek V4 Pro on OpenRouter was anomalous. The corrected price — $0.9553 input, $1.9105 output — matches the September level exactly. No pricing event occurred.
Georgia Tech and Emory Commit $1.175 Billion to Build Atlanta Into a National AI Hub
Georgia Tech ($575M+) and Emory University ($600M) announce joint AI investments targeting advanced computing, healthcare, law, ethics, and workforce development. The combined $1.175B is the largest dual-university AI commitment in the US South.