01OpenAI Ships GPT-6 Astra: 99.9% on ARC-AGI-3 and a Head-On Fight with Claude Fable [AI Models & Research]OpenAI's new flagship rolls out to Plus, Pro, Business and Enterprise users plus the API, priced at Fable-matching $10/$50 per million tokens. It scores 99.9% on ARC-AGI-3 — but via OpenAI's custom Provider Adapter harness that reuses opaque reasoning state; the default harness lands at 62.7%. It posts 100% on ExploitBench, yet Artificial Analysis still ranks Claude Fable 5.1 five points ahead on intelligence. → source
02Meta's Muse Spark 1.3 Ranks Third on the Intelligence Index — Fourth Release in Five Months [AI Models & Research]Meta's latest Muse Spark scores 62 on Artificial Analysis' Intelligence Index (max, partner preview) — behind only Claude Fable 5.1 and Opus 5 — while the shipping xhigh tier ties GPT-5.6 Sol. Gains concentrate in agentic and scientific work, and it claims DeepSWE's top spot at 75.4%. An open-weights version and a larger 'Watermelon' model are confirmed in the pipeline. → source
03Compile by Training Turns Plain-Language Specs into Local Neural Functions [AI Models & Research]Deng, Nie and Shieber propose compiling a natural-language specification into a reusable local neural function — distilling what you'd otherwise call a large remote model for into a small artifact you own. For recurring text tasks it cuts cost, latency and provider dependency in one move. → source
04When the LLM Judge Is the Instrument: Preregistered Audit Finds Unstable Measurements [AI Models & Research]LLM judges now gate training data, score generations and drive leaderboards — all resting on one rarely stated assumption: the same request to the same model name reads the same tomorrow. Two preregistered audits of black-box judges on shared endpoints found that assumption failing. → source
05Hugging Face's funes Gives Coding Agents a Durable Memory You Own [AI Tools & Ecosystem]funes turns the session logs your coding agents already produce into an indexed, retrievable memory layer — for Claude Code, Codex, pi and Hermes — with exact provenance, running locally. Its thesis: a memory is a dataset, not a service, so switching agents no longer means starting from zero. → source
06Abliteration.ai Turns Guardrail Removal into a Commercial API Business [AI Tools & Ecosystem]The startup hosts open-weight models with refusals stripped — including Z.ai's GLM-5.3 — queryable from a browser or API, aimed at offensive cyber, red-teaming and agent-testing work other models refuse to do. Abliteration has been an underground open-source practice for years; now it has a checkout page. → source
07Google's WeatherNext 3 Brings Hourly Satellite-Driven Forecasts to Search and Maps [AI Tools & Ecosystem]DeepMind's flagship weather AI now ingests real-time satellite data, refreshes hourly, resolves five times sharper than its predecessor, and adds clean-energy variables. It's live across Search, Gemini, Maps and Google Cloud — AI forecasting you already use without noticing. → source
08Nvidia Confirms the Deal Is Done: Hugging Face Sells for $12.9 Billion [AI Applications & Industry]Two weeks after reports of a $12.9B agreement that was 'not yet signed', Nvidia has confirmed the acquisition. Hugging Face's 3 million models, 1 million applications and 18 million developers — the distribution layer of open-weight AI — now sit inside the chipmaker. → source
09Accel in Talks to Lead $1B for Thinking Machines at a $40B Valuation [AI Applications & Industry]Mira Murati's startup would raise at $40B — below the $50B it sought last year — on a revenue run rate above $100M, an extraordinarily high multiple. Co-founder departures, including Lilian Weng back to OpenAI, shadow the round. → source
10Crusoe Raises $3B at $30B as Data-Center Money Keeps Compounding [AI Applications & Industry]Crusoe — the data-center developer behind builds for Meta, Microsoft and OpenAI — has raised $3B at a $30B valuation per Bloomberg, the latest sign that AI infrastructure financing now runs at sovereign-fund scale. → source
11Meta Will Pay to Watch How You Use Its New AI Model [AI Applications & Industry]Meta's new contributor tier pays users in exchange for visibility into how they use its latest AI model — training data by explicit incentive rather than silent default. It follows the internal employee-tracking program that was paused in June after wide criticism. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.