Anthropic's Claude Fable 5.1 ranks first on the index (score 57), edging out OpenAI's GPT 6 Astra, and scores 52.6% on Terminal Bench Science while cutting cost through cheaper cache reads.
Over 19 hours, 1,393 Hermes Agent subagents cleaned up a million-line Python codebase, shrinking it 34.4% and saving an estimated $2M in engineering time.
The next-generation general-purpose humanoid is designed to work near humans without physical safety barriers, ahead of the company's proposed SPAC merger. Agility reports more than $300M in multi-year customer orders, with early access in 2027.
An MLX runtime combining compressed KV execution, component residency and state preservation lets a laptop handle 200K-token contexts.
A prompt-only framework conditions answer commitment on an explicit assessment of the required information, cutting confident-but-wrong responses across three settings.
An empirical comparison of injecting new knowledge via the context window, latent representations or parameters, characterizing the accuracy-cost trade-offs of each route.
An interactive research workspace that continually improves its agents by turning researcher requests, feedback and execution evidence into tasks and evaluation rubrics.
A continuously running environment where failures propagate through memory, tools and environmental state, offering a safety regime beyond isolated model responses.
Meta will deploy its third-generation in-house chip in data centers in the first half of 2027, targeting more than a gigawatt of committed capacity over 12 months. Twelve test chips arrived from TSMC on September 1 and perform within 2-3% of simulation.
The SoC uses Arm C2-Ultra cores and two NPUs with improved on-device AI, alongside a 3nm 9600M for the broader flagship segment. First devices launch this quarter.
Priced at RMB 95,000 in China, the G1+ adds two neck degrees of freedom, much higher shoulder-and-waist and arm torque, a binocular vision system, head touch sensing and a six-microphone array.
The first IRON completed on the company's Guangzhou line walked off the floor autonomously. Mass production is targeted by the end of 2026, with deployments at its own stores and campuses.
The analysis finds roughly 20GW sits inside restricted local boundaries but only about 1,525MW is actually blocked, arguing that fears moratoriums are halting the US AI buildout are overstated.
AI Contact Hotline and agenthotline.ai let AI agents report misbehaving peers, following incidents of agent collusion.
Built by former TikTok product leads, the app uses an AI guide to generate pose suggestions for better portrait photos.
Meta launched a new tiered subscription bundling expanded AI usage across Facebook, Instagram and WhatsApp, plus business and creator tiers.
Shared, persistent cloud environments purpose-built for coding agents and human developers to collaborate.
Agent state lives in versionable, diffable files instead of a proprietary service.
The Rome-based cybersecurity firm secured the round led by Headline at a $1.7B valuation. Exein's Photon runtime provides kernel-level security for physical AI and IoT edge devices, and it says it protects over 2 billion connected devices while growing 400% year over year.
The innovation-intelligence platform counts 8,500 active users and clients including Leonardo, Poste Italiane and Banca Sella.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
