01Claude Haiku 5.5 Matches GPT-6 Luna's Pricing - and Anthropic Starts Paying Subscribers in API Credits [AI Models & Research]Anthropic's new small model exactly matches GPT-6 Luna at $0.10/$0.50 per million tokens up to 100K context and scores higher on benchmarks - though Willison flags a hidden price rise in a less generous tokenizer. Max and Team plans also gain monthly API credits, and Sonnet 5.5 cache reads are halved. → source
02Nemotron Takes Gold at IOI and IMO 2026 - and Beats the Top Human Score in Informatics [AI Models & Research]Starting from one Nemotron 3 family, NVIDIA's SFT, RL and generate-verify-refine recipes reached gold at both Olympiads: 535.4/600 at IOI - above the 498.27 top human score - and 30/42 at IMO. The models, data and recipes are open on Hugging Face. → source
03Liquid AI Opens d1: Single-Pass Decision Models for the Edge, From 600M to 3B [AI Models & Research]Two open decision models - d1-3B and d1-omni-600M - answer in a single forward pass instead of generating tokens. d1-3B tops a 4B rival on the mean of seven decision benchmarks, and the omni-600M reads text, images and audio, built for phones and other edge silicon. → source
04EngramEdit: Editing an LLM's Factual Memory Without Retraining It [AI Models & Research]A new arXiv paper decouples factual knowledge in conditional-memory architectures like DeepSeek Engram: n-gram lookups can carry edited facts without touching the base weights - a concrete path to models whose knowledge updates like a database row. → source
05Microsoft Puts Nvidia in the Surface: RTX Spark AI PCs and 'Execution Containers' for Agents [AI Tools & Ecosystem]The first Nvidia RTX Spark Windows machines land: a Surface Laptop Ultra from $2,600 and a $6,000 Dev Box. The bigger news is coming to all Windows 11 users - Execution Containers that sandbox third-party AI agents, plus up to $1,000 off for MacBook Pro trade-ins. → source
06ChatGPT Gets a Visual Brain: Intelligent UI Rolls Out With a New GPT-6 Model [AI Tools & Ecosystem]OpenAI is reworking the text-first chat into an interactive surface backed by a new GPT-6 model: tappable buttons, custom calculators, editable charts and instant diagrams. It ships to Pro, Plus, Business and Enterprise first, with Free and Go users a day later. → source
07Google Opens SynthID Verification to Everyone - 1 Million Checks a Day [AI Tools & Ecosystem]Anyone can now upload an image, video or audio clip and check it for SynthID's invisible watermark. OpenAI, NVIDIA and Kakao already support the standard and Apple is said to be next - provenance checking is becoming a default platform feature rather than a lab demo. → source
08Google's Playground Turns Text Prompts Into Browser Games [AI Tools & Ecosystem]Google Labs' Playground generates playable 2D and 3D browser games from text prompts, with uploaded images turned into game assets, a publish gallery with leaderboards, and Unity Spark integration on the roadmap - a weekly token budget meters generation. → source
09Nous Research Confirms $1.5B Valuation - and Takes the Hermes Agent to Business [AI Applications & Industry]A $90M Series B led by Robot Ventures puts the open-source Hermes agent's maker at $1.5B, with 24M clones and an estimated 2.5% of global AI token usage. The push: 'Hermes for Businesses' for private multi-step enterprise agents, targeting $100M run rate by year-end off $36M in September. → source
10Meta's New AI Drills Into Ads That Secretly Lead to Child Abuse Material [AI Applications & Industry]Meta acted on 33.2 million pieces of child sexual exploitation content in H1 2026 - 97% found by its systems first - and now targets 'signposting' ads that look harmless but route users underground, judging ads by where they send people. A red-teaming agent probes its own defenses. → source
11Tony Fadell's Verdict on Gen-1 AI Gadgets: Most of Us Have Never Had an Assistant [AI Applications & Industry]The iPod's creator lines up the R1, Humane Pin and Limitless pendant as failures that solved no real need - fewer than 0.01% of people have ever had a helper, so the pitch never had an audience. The agent era's scarce resource is trust, and companies keep trying to ship it in a device. → source
12Healthleap Raises $38M to Read Hospital Notes and Flag Patients Doctors Might Miss [AI Applications & Industry]Sequoia, First Round and Hummingbird back a clinical-notes screener live in 50+ hospitals that reads free-text notes for signs of malnutrition and delirium - priced on outcome, with Penn Medicine citing $23.8 million a year in validated financial impact. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.