Live Feed
Muse Spark 1.2's 2.9-Point Terminal-Bench Gain Is a Harness Delta, Not a Model Delta
Meta co-trained Muse Spark 1.2 with Muse Code and then measured it inside Muse Code, while measuring 1.1 in a different harness. The claimed gain is smaller than the mean absolute harness delta on the verified board — and neither Muse Spark 1.2 nor its named competitors appear on the independent leaderboard at all.
GLM-5.3 Enters GDPval-AA at #3 with ELO 1769, Above Grok 4.6
Z AI's GLM-5.3 (max) debuted on the GDPval-AA v2 leaderboard this week at ELO 1769, placing third overall behind two Claude Opus 5 variants and ahead of Grok 4.6 High at 1747. A Chinese closed model now holds a top-four seat in Artificial Analysis's flagship agentic benchmark.
Frontier Reasoning Models Score Below 16% on Chain-of-Thought Controllability, OpenAI Says That's Good
OpenAI tested 13 frontier reasoning models and found that none can reliably comply with instructions about how to reason. Controllability scores range from 0.1% to 15.4%. The company is reframing the limitation as a safety property.
AI Boosted Homework Scores 30%, Then Crashed Exam Performance: 27,000-Student Study
A large-scale study covering 27,000 middle and high school students in China found AI tools significantly improved homework performance in the short term, then caused serious exam score declines. High achievers were hit hardest.
GPT-5.6-Sol Ignores Sub-Agent Teardown Instructions, Costs Developer $500 in One Run
Multi-agent pipelines on GPT-5.6-Sol at very high effort are reporting a specific failure: the orchestrator ignores instructions to close sub-agents between phases, burning over $500 per session. The community fix is treating the model like a new hire who needs explicit, documented constraints.
Wispr Raises $280M at $2B: Menlo Ventures Bets the Text Box Is AI's Weakest Link
Menlo Ventures led a $280M Series B into voice dictation startup Wispr at a $2B valuation. The thesis is specific: AI has solved the model layer, and the interface is now the bottleneck. Text input is next to break.
DeepSeek Adds Vision to V4 Flash for Free: V4-Flash-Vision-Exp Processes Images and Screenshots
DeepSeek's experimental multimodal extension of V4 Flash launched Friday with image and screenshot interpretation at no charge on its API. Text performance matches the base model.
Gemma Crosses 1 Billion Downloads: NASA Is Running It in Orbit, Developers Have Published 100,000 Variants
Google DeepMind's open model family hit one billion downloads two years after launch. Teams at NASA, Satlyt, and Starcloud operate Gemma in orbit for onboard image analysis. Over 100,000 variants published on Hugging Face.
OpenAI Codex Adds Multi-Agent V2: Sol Can Now Delegate Tasks to Luna Subagents to Cut Credit Burn
OpenAI shipped Multi-Agent V2 to Codex this week, letting GPT-5.6 Sol spawn Luna agents for bounded subtasks instead of resolving everything at Sol's token rate. The orchestrator can delegate to any supported model, not just Luna.
Microsoft MAI-Image-2.6 Debuts at #2 on Arena Image Edit, Ahead of Google, Meta, and xAI
Microsoft AI Image 2.6 entered Arena's Image Edit leaderboard at number two on August 18, overtaking models from Google, Meta, and xAI. The prior generation, MAI-Image-2.5, debuted at #3 in May.
Amazon Is Buying Rare Books in Bulk, Stripping Their Spines, and Pulping the Rest
Amazon workers at a Nevada warehouse strip the spines off rare physical books received in bulk, scan the pages, and send the physical copies for pulping. The practice mirrors moves by Anthropic and Google and has alarmed rare book collectors and archivists.
Ox Alpha Appears on OpenRouter Free: Stealth Frontier Model Identified as GLM Architecture
A model called Ox Alpha surfaced on OpenRouter under a stealth namespace with a 1M-token context window, free pricing, and agentic coding positioning. Community analysis links it to Zhipu AI's GLM architecture.
Open Weights Match Frontier on Agentic Coding but Trail by 12 Points on Mathematics, LiveBench Shows
Smaug-Agentic and Kimi K3 score 64.6 and 62.2 on LiveBench agentic coding — matching or beating Claude Fable 5 — but score only 83.9 and 84.4 on mathematics while frontier closed models cluster at 95.7 to 96.2. The gap reveals what agentic fine-tuning can and cannot fix.
Claude Can Now Send Emails in Gmail Without Per-Action Approval
Anthropic has added autonomous email sending to Claude's Gmail integration. Claude can now dispatch emails on a user's behalf without requesting confirmation for each action — the first major communication channel where Claude operates with unsanctioned send authority.
Anthropic Investors Target $2 Trillion at October IPO — Would Be Largest Ever
Six Anthropic backers expect the company to price its October IPO at $2 trillion or more, more than double the $965 billion valuation at which it filed its S-1. Annualized revenue is tracking toward $100-120 billion by year end, up tenfold from the $47 billion figure reported in May.