AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →OpenRouter launches new Auto and Auto Code routers with up to 4x performance gains
The new routers weigh eleven factors per request and handle over 2.5T tokens monthly, with Auto Code matching Claude Opus 5.5 performance at 74% lower cost.
NVIDIA LPR Introduces VFlash: A Generalist Decoding Framework for Physical AI Visual Planning
The framework uses parallel visual primitive drafting to reduce 3D displacement error by up to 61% and accelerate driving inference by up to 49x compared to autoregressive decoding.
Cursor enables live app deployment and sharing via link
The feature automates database setup and runs deterministic and agentic security checks before deployment.
Dwarkesh Patel's Blood on the Clocktower benchmark shows GPT-6 Astra outperforms Claude Opus 4.6 in strategic deception
In 386 games of the social deduction game Blood on the Clocktower, GPT-6 Astra achieved a 92% win rate against Claude Opus 4.6, with OpenAI models showing superior critical evaluation of evidence.
Meta rolls out voice and expressive avatars in the Muse app
The feature is available on mobile, web, and Mac in the US, with AI Glasses and Muse Charm support coming soon.
Norm Law Designs AI-Native Firm Structure to Unlock Nonlinear Productivity
The firm argues that existing human-centric structures fail to capture AI value, advocating for AI-native designs where agents and humans co-evolve to drive nonlinear productivity gains.
Ornn Explains Primary Compute Market Structures: Reserved, On-Demand, and Spot
The article uses a hypothetical 1,024 NVIDIA B300 cluster to illustrate how reserved, on-demand, and spot pricing models function in the primary compute market.
PickOpenAI launches Images 2.6 with 50% faster generation and GIF support
The update is available across all ChatGPT tiers and the API, with Flare and Sunburst models supporting SVG exports and a new complexity parameter.
PickCat Wu joins OpenAI to lead product management for Core Products
Wu cites OpenAI's reach of over a billion users and plans to build agents that keep users in control.
Devin Review 2.0 Cuts PR Review Costs by 82% and Becomes Default Tool
The update makes Devin Review the default for PRs, with a median latency of 94 seconds and new features for incremental reviews and custom checks.
METR Finds Anthropic's AI R&D Acceleration at 1.9x Historical Rate
Self-reported AI uplift at Anthropic rose from 1.2x to 12x, with no evidence of algorithmic or compute bottlenecks in the next 12 months.
Study: Frontier LLMs maintain accuracy with short contexts via frequent compaction
The disparity in generation cost between later and earlier tokens on frontier open-weight models can reach 7x.
Google DeepMind open-sources AlphaEvolve code and verification notebooks for new math results
The repository includes a minimal implementation of AlphaEvolve, data for a new upper bound on the matrix multiplication exponent, and verification code for 15 mathematical problems.
Axiom releases Lean formalization of the 7,500-page Stacks Project
The 16 million-line library is the largest integrated formalized mathematics collection, enabling direct import of algebraic geometry definitions and theorems.
Cognition Releases SWE-2 Flash, a 276B Coding Model Outperforming Larger American Open-Weights
The 276B-A12B model achieves 20.0% on Terminal-Bench 4.0 and 24.0% on FrontierCode 2.0, running at 200 tokens per second.
Tencent Releases Hy4, Ranking Among World's Leading Open-Source AI Models with New Visual and Computer-Use Capabilities
The model achieves an average WorkBuddy score of 83.18 and reduces hallucination rates from 12.95% to 7.2% while maintaining pricing of US$0.90 per million input tokens.
Perplexity Decisions API adds web grounding, boosting accuracy from 51% to 86% on PWS-Decisions
The feature searches the web before answering, raising accuracy on current-fact questions from 51% to 86% while keeping median latency under 400 ms.
Together AI Releases TKCC Compiler for Continual AI Inference Optimization
The compiler uses persistent engines to reduce startup time from 467 seconds to 9.0 seconds, enabling real-time adaptation to traffic and hardware changes.
Meta Acquires AI Coding Startup Conductor
Conductor Cloud will go read-only on November 13 and shut down on November 20, while the local version receives updates until early 2027.
Pangram releases Image 1.0 and Flash 1.0 for AI image and video detection
Pangram releases Image 1.0 and Flash 1.0, with the video model achieving a 29-point lead over competitors on the DF26 benchmark.